mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2026-09-19 01:05:09 +02:00
Merge branch 'upstream' into concedo_experimental
# Conflicts: # .github/workflows/ui-build.yml # .github/workflows/ui-publish.yml # .github/workflows/ui-self-hosted.yml # CMakeLists.txt # app/CMakeLists.txt # app/llama.cpp # common/arg.cpp # ggml/CMakeLists.txt # ggml/src/ggml-hexagon/ggml-hexagon.cpp # ggml/src/ggml-hexagon/htp/htp-ops.h # ggml/src/ggml-hexagon/htp/main.c # ggml/src/ggml-hexagon/htp/unary-ops.c # ggml/src/ggml-opencl/ggml-opencl.cpp # scripts/snapdragon/ggml-hexagon-profile.py # scripts/sync-ggml.last # src/CMakeLists.txt # tests/test-llama-archs.cpp # tools/cli/README.md # tools/completion/README.md # tools/mtmd/CMakeLists.txt # tools/mtmd/tests/test-deepseek-ocr.py # tools/server/README.md
This commit is contained in:
Concedo
@@ -0,0 +1,43 @@
|
||||
name: UI Build (self-hosted)
|
||||
|
||||
on:
|
||||
workflow_call:
|
||||
|
||||
jobs:
|
||||
build:
|
||||
runs-on: [self-hosted, fast]
|
||||
env:
|
||||
BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
|
||||
|
||||
steps:
|
||||
- name: Checkout code
|
||||
uses: actions/checkout@v6
|
||||
|
||||
- name: Setup Node.js
|
||||
uses: actions/setup-node@v6
|
||||
with:
|
||||
node-version: "24"
|
||||
cache: "npm"
|
||||
cache-dependency-path: "tools/ui/package-lock.json"
|
||||
|
||||
- name: Install dependencies
|
||||
run: npm ci
|
||||
working-directory: tools/ui
|
||||
|
||||
- name: Build application
|
||||
run: npm run build
|
||||
working-directory: tools/ui
|
||||
|
||||
- name: Generate checksums
|
||||
run: |
|
||||
cd tools/ui/dist
|
||||
for f in *; do
|
||||
sha256sum "$f" | awk '{print $1, $2}' >> checksums.txt
|
||||
done
|
||||
|
||||
- name: Upload built UI
|
||||
uses: actions/upload-artifact@v6
|
||||
with:
|
||||
name: ui-build
|
||||
path: tools/ui/dist/
|
||||
retention-days: 1
|
||||
+41
-46
@@ -51,9 +51,6 @@
|
||||
|
||||
#define LLAMA_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
|
||||
|
||||
extern const char * LICENSES[];
|
||||
const char * LICENSES[] = {};
|
||||
|
||||
using json = nlohmann::ordered_json;
|
||||
using namespace common_arg_utils;
|
||||
|
||||
@@ -344,9 +341,7 @@ struct handle_model_result {
|
||||
};
|
||||
|
||||
static handle_model_result common_params_handle_model(struct common_params_model & model,
|
||||
const std::string & bearer_token,
|
||||
bool offline,
|
||||
bool search_mtp = false) {
|
||||
const common_download_opts & opts) {
|
||||
handle_model_result result;
|
||||
|
||||
if (!model.docker_repo.empty()) {
|
||||
@@ -358,10 +353,9 @@ static handle_model_result common_params_handle_model(struct common_params_model
|
||||
model.hf_file = model.path;
|
||||
model.path = "";
|
||||
}
|
||||
common_download_opts opts;
|
||||
opts.bearer_token = bearer_token;
|
||||
opts.offline = offline;
|
||||
auto download_result = common_download_model(model, opts, true, search_mtp);
|
||||
common_download_opts hf_opts = opts;
|
||||
hf_opts.download_mmproj = true; // also look for mmproj when downloading hf model
|
||||
auto download_result = common_download_model(model, hf_opts);
|
||||
|
||||
if (download_result.model_path.empty()) {
|
||||
throw std::runtime_error("failed to download model from Hugging Face");
|
||||
@@ -386,9 +380,6 @@ static handle_model_result common_params_handle_model(struct common_params_model
|
||||
model.path = fs_get_cache_file(string_split<std::string>(f, '/').back());
|
||||
}
|
||||
|
||||
common_download_opts opts;
|
||||
opts.bearer_token = bearer_token;
|
||||
opts.offline = offline;
|
||||
auto download_result = common_download_model(model, opts);
|
||||
if (download_result.model_path.empty()) {
|
||||
throw std::runtime_error("failed to download model from " + model.url);
|
||||
@@ -445,35 +436,49 @@ static bool parse_bool_value(const std::string & value) {
|
||||
// CLI argument parsing functions
|
||||
//
|
||||
|
||||
void common_params_handle_models(common_params & params, llama_example curr_ex) {
|
||||
bool common_params_handle_models(common_params & params, llama_example curr_ex) {
|
||||
const bool spec_type_draft_mtp = std::find(params.speculative.types.begin(),
|
||||
params.speculative.types.end(),
|
||||
COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
|
||||
|
||||
auto res = common_params_handle_model(params.model, params.hf_token, params.offline, spec_type_draft_mtp);
|
||||
if (params.no_mmproj) {
|
||||
params.mmproj = {};
|
||||
} else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
|
||||
// optionally, handle mmproj model when -hf is specified
|
||||
params.mmproj = res.mmproj;
|
||||
}
|
||||
// only download mmproj if the current example is using it
|
||||
for (const auto & ex : mmproj_examples) {
|
||||
if (curr_ex == ex) {
|
||||
common_params_handle_model(params.mmproj, params.hf_token, params.offline);
|
||||
break;
|
||||
common_download_opts opts;
|
||||
opts.bearer_token = params.hf_token;
|
||||
opts.offline = params.offline;
|
||||
opts.skip_download = params.skip_download;
|
||||
opts.download_mtp = spec_type_draft_mtp;
|
||||
|
||||
try {
|
||||
auto res = common_params_handle_model(params.model, opts);
|
||||
if (params.no_mmproj) {
|
||||
params.mmproj = {};
|
||||
} else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
|
||||
// optionally, handle mmproj model when -hf is specified
|
||||
params.mmproj = res.mmproj;
|
||||
}
|
||||
// only download mmproj if the current example is using it
|
||||
for (const auto & ex : mmproj_examples) {
|
||||
if (curr_ex == ex) {
|
||||
common_params_handle_model(params.mmproj, opts);
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
// when --spec-type mtp is set and no draft model was provided explicitly,
|
||||
// fall back to the MTP head discovered alongside the -hf model
|
||||
if (spec_type_draft_mtp && res.found_mtp &&
|
||||
params.speculative.draft.mparams.path.empty() &&
|
||||
params.speculative.draft.mparams.hf_repo.empty() &&
|
||||
params.speculative.draft.mparams.url.empty()) {
|
||||
params.speculative.draft.mparams.path = res.mtp.path;
|
||||
}
|
||||
common_params_handle_model(params.speculative.draft.mparams, opts);
|
||||
common_params_handle_model(params.vocoder.model, opts);
|
||||
return true;
|
||||
} catch (const common_skip_download_exception &) {
|
||||
return false;
|
||||
} catch (const std::exception &) {
|
||||
throw;
|
||||
}
|
||||
// when --spec-type mtp is set and no draft model was provided explicitly,
|
||||
// fall back to the MTP head discovered alongside the -hf model
|
||||
if (spec_type_draft_mtp && res.found_mtp &&
|
||||
params.speculative.draft.mparams.path.empty() &&
|
||||
params.speculative.draft.mparams.hf_repo.empty() &&
|
||||
params.speculative.draft.mparams.url.empty()) {
|
||||
params.speculative.draft.mparams.path = res.mtp.path;
|
||||
}
|
||||
common_params_handle_model(params.speculative.draft.mparams, params.hf_token, params.offline);
|
||||
common_params_handle_model(params.vocoder.model, params.hf_token, params.offline);
|
||||
}
|
||||
|
||||
static bool common_params_parse_ex(int argc, char ** argv, common_params_context & ctx_arg) {
|
||||
@@ -1093,16 +1098,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
|
||||
exit(0);
|
||||
}
|
||||
));
|
||||
add_opt(common_arg(
|
||||
{"--license"},
|
||||
"show source code license and dependencies",
|
||||
[](common_params &) {
|
||||
for (int i = 0; LICENSES[i]; ++i) {
|
||||
printf("%s\n", LICENSES[i]);
|
||||
}
|
||||
exit(0);
|
||||
}
|
||||
));
|
||||
add_opt(common_arg(
|
||||
{"-cl", "--cache-list"},
|
||||
"show list of models in cache",
|
||||
|
||||
+5
-2
@@ -129,8 +129,11 @@ bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<com
|
||||
// see: https://github.com/ggml-org/llama.cpp/issues/18163
|
||||
void common_params_add_preset_options(std::vector<common_arg> & args);
|
||||
|
||||
// Populate model paths (main model, mmproj, etc) from -hf if necessary
|
||||
void common_params_handle_models(common_params & params, llama_example curr_ex);
|
||||
// populate model paths (main model, mmproj, etc) from -hf if necessary
|
||||
// return true if the model is ready to use
|
||||
// throw an exception if there is an error that prevents the model from being used (e.g. network error, model not found, etc)
|
||||
// if params.skip_download is true, no downloads will be attempted. return false if the model is invalid or missing (e.g. ETag check failed)
|
||||
bool common_params_handle_models(common_params & params, llama_example curr_ex);
|
||||
|
||||
// initialize argument parser context - used by test-arg-parser and preset
|
||||
common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
|
||||
|
||||
+3
-2
@@ -480,7 +480,7 @@ struct common_params {
|
||||
|
||||
std::set<std::string> model_alias; // model aliases // NOLINT
|
||||
std::set<std::string> model_tags; // model tags (informational, not used for routing) // NOLINT
|
||||
std::string hf_token = ""; // HF token // NOLINT
|
||||
std::string hf_token = ""; // HF token (aka bearer token) // NOLINT
|
||||
std::string prompt = ""; // NOLINT
|
||||
std::string system_prompt = ""; // NOLINT
|
||||
std::string prompt_file = ""; // store the external prompt file name // NOLINT
|
||||
@@ -508,6 +508,7 @@ struct common_params {
|
||||
int32_t control_vector_layer_start = -1; // layer range for control vector
|
||||
int32_t control_vector_layer_end = -1; // layer range for control vector
|
||||
bool offline = false;
|
||||
bool skip_download = false; // skip model file downloading
|
||||
|
||||
int32_t ppl_stride = 0; // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
|
||||
int32_t ppl_output_type = 0; // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
|
||||
@@ -588,7 +589,7 @@ struct common_params {
|
||||
// server params
|
||||
int32_t port = 8080; // server listens on this network port
|
||||
bool reuse_port = false; // allow multiple sockets to bind to the same port
|
||||
int32_t timeout_read = 600; // http read timeout in seconds
|
||||
int32_t timeout_read = 3600; // http read timeout in seconds
|
||||
int32_t timeout_write = timeout_read; // http write timeout in seconds
|
||||
int32_t n_threads_http = -1; // number of threads to process HTTP requests (TODO: support threadpool)
|
||||
int32_t n_cache_reuse = 0; // min chunk size to reuse from the cache via KV shifting
|
||||
|
||||
+19
-7
@@ -295,6 +295,10 @@ static int common_download_file_single_online(const std::string & url,
|
||||
|
||||
const bool file_exists = std::filesystem::exists(path);
|
||||
|
||||
if (!file_exists && opts.skip_download) {
|
||||
return -2; // file is missing and download is disabled
|
||||
}
|
||||
|
||||
if (file_exists && skip_etag) {
|
||||
LOG_DBG("%s: using cached file: %s\n", __func__, path.c_str());
|
||||
return 304; // 304 Not Modified - fake cached response
|
||||
@@ -360,6 +364,10 @@ static int common_download_file_single_online(const std::string & url,
|
||||
LOG_DBG("%s: using cached file (same etag): %s\n", __func__, path.c_str());
|
||||
return 304; // 304 Not Modified - fake cached response
|
||||
}
|
||||
// pass this point, the file exists but is different from the server version, so we need to redownload it
|
||||
if (opts.skip_download) {
|
||||
return -2; // special code to indicate that the download was skipped due to etag mismatch
|
||||
}
|
||||
if (remove(path.c_str()) != 0) {
|
||||
LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
|
||||
return -1;
|
||||
@@ -778,13 +786,13 @@ static std::vector<download_task> get_url_tasks(const common_params_model & mode
|
||||
}
|
||||
|
||||
common_download_model_result common_download_model(const common_params_model & model,
|
||||
const common_download_opts & opts,
|
||||
bool download_mmproj,
|
||||
bool download_mtp) {
|
||||
const common_download_opts & opts) {
|
||||
common_download_model_result result;
|
||||
std::vector<download_task> tasks;
|
||||
hf_plan hf;
|
||||
|
||||
bool download_mmproj = opts.download_mmproj;
|
||||
bool download_mtp = opts.download_mtp;
|
||||
bool is_hf = !model.hf_repo.empty();
|
||||
|
||||
if (is_hf) {
|
||||
@@ -809,18 +817,22 @@ common_download_model_result common_download_model(const common_params_model &
|
||||
return result;
|
||||
}
|
||||
|
||||
std::vector<std::future<bool>> futures;
|
||||
std::vector<std::future<int>> futures;
|
||||
for (const auto & task : tasks) {
|
||||
futures.push_back(std::async(std::launch::async,
|
||||
[&task, &opts, is_hf]() {
|
||||
int status = common_download_file_single(task.url, task.path, opts, is_hf);
|
||||
return is_http_status_ok(status);
|
||||
return common_download_file_single(task.url, task.path, opts, is_hf);
|
||||
}
|
||||
));
|
||||
}
|
||||
|
||||
for (auto & f : futures) {
|
||||
if (!f.get()) {
|
||||
int status = f.get();
|
||||
if (status == -2 && opts.skip_download) {
|
||||
throw common_skip_download_exception();
|
||||
}
|
||||
bool is_ok = is_http_status_ok(status);
|
||||
if (!is_ok) {
|
||||
return {};
|
||||
}
|
||||
}
|
||||
|
||||
+10
-3
@@ -52,6 +52,9 @@ struct common_download_opts {
|
||||
std::string bearer_token;
|
||||
common_header_list headers;
|
||||
bool offline = false;
|
||||
bool skip_download = false; // if true, only validation is performed, common_skip_download_exception may be thrown if the file is missing or invalid
|
||||
bool download_mmproj = false;
|
||||
bool download_mtp = false;
|
||||
common_download_callback * callback = nullptr;
|
||||
};
|
||||
|
||||
@@ -62,6 +65,11 @@ struct common_download_model_result {
|
||||
std::string mtp_path;
|
||||
};
|
||||
|
||||
// throw if the file is missing or invalid (e.g. ETag check failed)
|
||||
struct common_skip_download_exception : public std::runtime_error {
|
||||
common_skip_download_exception() : std::runtime_error("skip download") {}
|
||||
};
|
||||
|
||||
// Download model from HuggingFace repo or URL
|
||||
//
|
||||
// input (via model struct):
|
||||
@@ -89,9 +97,7 @@ struct common_download_model_result {
|
||||
// returns result with model_path, mmproj_path and mtp_path (empty when not found / on failure)
|
||||
common_download_model_result common_download_model(
|
||||
const common_params_model & model,
|
||||
const common_download_opts & opts = {},
|
||||
bool download_mmproj = false,
|
||||
bool download_mtp = false
|
||||
const common_download_opts & opts = {}
|
||||
);
|
||||
|
||||
// returns list of cached models
|
||||
@@ -99,6 +105,7 @@ std::vector<common_cached_model_info> common_list_cached_models();
|
||||
|
||||
// download single file from url to local path
|
||||
// returns status code or -1 on error
|
||||
// returns -2 if the download was skipped due to ETag mismatch (file outdated, skip_download=true)
|
||||
// skip_etag: if true, don't read/write .etag files (for HF cache where filename is the hash)
|
||||
int common_download_file_single(const std::string & url,
|
||||
const std::string & path,
|
||||
|
||||
@@ -1,5 +1,7 @@
|
||||
#include "ngram-mod.h"
|
||||
|
||||
#include <algorithm>
|
||||
|
||||
//
|
||||
// common_ngram_mod
|
||||
//
|
||||
|
||||
@@ -47,6 +47,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
|
||||
"DeepseekForCausalLM": "deepseek",
|
||||
"DeepseekV2ForCausalLM": "deepseek",
|
||||
"DeepseekV3ForCausalLM": "deepseek",
|
||||
"DeepseekV32ForCausalLM": "deepseek",
|
||||
"DistilBertForMaskedLM": "bert",
|
||||
"DistilBertForSequenceClassification": "bert",
|
||||
"DistilBertModel": "bert",
|
||||
@@ -236,6 +237,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
|
||||
MMPROJ_MODEL_MAP: dict[str, str] = {
|
||||
"AudioFlamingo3ForConditionalGeneration": "ultravox",
|
||||
"CogVLMForCausalLM": "cogvlm",
|
||||
"DeepseekOCR2ForCausalLM": "deepseek",
|
||||
"DeepseekOCRForCausalLM": "deepseek",
|
||||
"DotsOCRForCausalLM": "dotsocr",
|
||||
"Gemma3ForConditionalGeneration": "gemma",
|
||||
|
||||
+3
-1
@@ -915,6 +915,8 @@ class ModelBase:
|
||||
gguf.MODEL_TENSOR.SSM_CONV1D_Q,
|
||||
gguf.MODEL_TENSOR.SSM_CONV1D_K,
|
||||
gguf.MODEL_TENSOR.SSM_CONV1D_V,
|
||||
# DSA indexer weights should be F32
|
||||
gguf.MODEL_TENSOR.INDEXER_PROJ,
|
||||
)
|
||||
)
|
||||
or new_name[-7:] not in (".weight", ".lora_a", ".lora_b")
|
||||
@@ -1138,7 +1140,7 @@ class TextModel(ModelBase):
|
||||
# Skip multimodal tensors
|
||||
if name.startswith(("mlp", "vit.", "vpm.", "siglip2.", "conformer.", "merger.", "resampler.", "sound_encoder.", "sound_projection.", "speech_embeddings.")) \
|
||||
or "visual." in name or "vision." in name or "audio." in name or "talker." in name \
|
||||
or "vision_" in name or "audio_" in name or "sam_model" in name \
|
||||
or "vision_" in name or "audio_" in name \
|
||||
or "token2wav." in name or "code2wav." in name \
|
||||
or "projector." in name or "pre_mm_projector_norm" in name \
|
||||
or "image_newline" in name or "view_seperator" in name \
|
||||
|
||||
+85
-12
@@ -16,10 +16,14 @@ from .qwen import QwenModel
|
||||
|
||||
@ModelBase.register("DeepseekOCRForCausalLM")
|
||||
class DeepseekOCRVisionModel(MmprojModel):
|
||||
def __init__(self, *args, **kwargs):
|
||||
super().__init__(*args, **kwargs)
|
||||
self.clip_projector_type = gguf.VisionProjectorType.DEEPSEEKOCR
|
||||
|
||||
def set_gguf_parameters(self):
|
||||
super().set_gguf_parameters()
|
||||
hparams = self.hparams
|
||||
self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.DEEPSEEKOCR)
|
||||
self.gguf_writer.add_clip_projector_type(self.clip_projector_type)
|
||||
# default values below are taken from HF tranformers code
|
||||
self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("layer_norm_eps", 1e-6))
|
||||
self.gguf_writer.add_vision_use_gelu(True)
|
||||
@@ -49,22 +53,27 @@ class DeepseekOCRVisionModel(MmprojModel):
|
||||
raise ValueError("DeepseekOCR model requires 'vision_config' in the model configuration, but it was not found")
|
||||
|
||||
vision_config['sam'] = vision_config['width']['sam_vit_b']
|
||||
vision_config.update(vision_config['width']['clip-l-14-224'])
|
||||
vision_config['hidden_size'] = vision_config['width']
|
||||
vision_config['num_heads'] = vision_config['heads']
|
||||
vision_config['intermediate_size'] = vision_config['heads'] * 4
|
||||
if vision_config['width'].get('clip-l-14-224') is not None:
|
||||
vision_config.update(vision_config['width']['clip-l-14-224'])
|
||||
if isinstance(vision_config['width'], int):
|
||||
vision_config['hidden_size'] = vision_config['width']
|
||||
if vision_config.get('heads') is not None:
|
||||
vision_config['num_heads'] = vision_config['heads']
|
||||
vision_config['intermediate_size'] = vision_config['heads'] * 4
|
||||
|
||||
return vision_config
|
||||
|
||||
def tensor_force_quant(self, name, new_name, bid, n_dims):
|
||||
if ".embeddings." in name or 'pos_embed' in name:
|
||||
return gguf.GGMLQuantizationType.F32
|
||||
if ".rel_pos_h" in name or '.rel_pos_w' in name:
|
||||
return gguf.GGMLQuantizationType.F32
|
||||
if ".neck." in name or ".net_" in name:
|
||||
return gguf.GGMLQuantizationType.F32
|
||||
for nq_name in ('.embeddings.', 'pos_embed', '.rel_pos_h', '.rel_pos_w', '.neck.', '.net_'):
|
||||
if nq_name in name:
|
||||
return gguf.GGMLQuantizationType.F32
|
||||
return super().tensor_force_quant(name, new_name, bid, n_dims)
|
||||
|
||||
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
|
||||
if name.endswith("view_seperator"):
|
||||
data_torch = data_torch.unsqueeze(0)
|
||||
yield from super().modify_tensors(data_torch, name, bid)
|
||||
|
||||
@classmethod
|
||||
def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
|
||||
name, gen = item
|
||||
@@ -81,6 +90,33 @@ class DeepseekOCRVisionModel(MmprojModel):
|
||||
return super().filter_tensors((name, gen))
|
||||
|
||||
|
||||
@ModelBase.register("DeepseekOCR2ForCausalLM")
|
||||
class DeepseekOCR2VisionModel(DeepseekOCRVisionModel):
|
||||
def __init__(self, *args, **kwargs):
|
||||
super().__init__(*args, **kwargs)
|
||||
self.clip_projector_type = gguf.VisionProjectorType.DEEPSEEKOCR2
|
||||
|
||||
def set_gguf_parameters(self):
|
||||
# the vision tower's qwen2 encoder is built from fixed defaults,
|
||||
# see build_qwen2_decoder_as_encoder() in deepencoderv2.py
|
||||
if self.hparams.get("patch_size") is None:
|
||||
self.hparams["patch_size"] = 16
|
||||
if self.hparams.get("intermediate_size") is None:
|
||||
self.hparams["intermediate_size"] = 4864
|
||||
if self.hparams.get("num_attention_heads") is None:
|
||||
self.hparams["num_attention_heads"] = 14
|
||||
super().set_gguf_parameters()
|
||||
# qwen2 encoder is GQA: 14 Q heads, 2 KV heads
|
||||
self.gguf_writer.add_vision_head_count_kv(2)
|
||||
|
||||
def get_vision_config(self) -> dict[str, Any]:
|
||||
vision_config = super().get_vision_config()
|
||||
vision_config['hidden_size'] = vision_config['width']['qwen2-0-5b']['dim']
|
||||
if vision_config.get('layers') is None:
|
||||
vision_config['layers'] = 24
|
||||
return vision_config
|
||||
|
||||
|
||||
@ModelBase.register("DeepseekForCausalLM")
|
||||
class DeepseekModel(TextModel):
|
||||
model_arch = gguf.MODEL_ARCH.DEEPSEEK
|
||||
@@ -188,13 +224,21 @@ class DeepseekV2Model(TextModel):
|
||||
self.origin_hf_arch = hparams.get('architectures', [None])[0]
|
||||
|
||||
# special handling for Deepseek OCR
|
||||
if self.origin_hf_arch == "DeepseekOCRForCausalLM":
|
||||
if self.origin_hf_arch in ("DeepseekOCRForCausalLM", "DeepseekOCR2ForCausalLM"):
|
||||
self.model_arch = gguf.MODEL_ARCH.DEEPSEEK2OCR
|
||||
self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[self.model_arch]
|
||||
self.gguf_writer.add_architecture()
|
||||
# default jinja template
|
||||
self.gguf_writer.add_chat_template("{% for m in messages %}{{m['content']}}{% endfor %}")
|
||||
|
||||
@classmethod
|
||||
def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
|
||||
name, _ = item
|
||||
# DeepSeek-OCR vision encoder (SAM + DeepSeek-OCR-2 qwen2 tower)
|
||||
if "sam_model" in name or "qwen2_model" in name:
|
||||
return None
|
||||
return super().filter_tensors(item)
|
||||
|
||||
def set_vocab(self):
|
||||
try:
|
||||
self._set_vocab_gpt2()
|
||||
@@ -386,3 +430,32 @@ class DeepseekV2Model(TextModel):
|
||||
experts = [k for d in self._experts for k in d.keys()]
|
||||
if len(experts) > 0:
|
||||
raise ValueError(f"Unprocessed experts: {experts}")
|
||||
|
||||
|
||||
@ModelBase.register("DeepseekV32ForCausalLM")
|
||||
class DeepseekV32Model(DeepseekV2Model):
|
||||
model_arch = gguf.MODEL_ARCH.DEEPSEEK32
|
||||
skip_mtp = False
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
super().__init__(*args, **kwargs)
|
||||
self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
|
||||
self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
|
||||
|
||||
def set_vocab(self):
|
||||
from transformers import AutoTokenizer
|
||||
tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
|
||||
assert getattr(tokenizer, "add_bos_token", False), "Change value of add_bos_token to true in tokenizer_config.json file."
|
||||
self._set_vocab_gpt2()
|
||||
|
||||
def set_gguf_parameters(self):
|
||||
super().set_gguf_parameters()
|
||||
|
||||
# NextN/MTP prediction layers
|
||||
if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
|
||||
self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
|
||||
|
||||
# DSA indexer parameters
|
||||
self.gguf_writer.add_indexer_head_count(self.hparams["index_n_heads"])
|
||||
self.gguf_writer.add_indexer_key_length(self.hparams["index_head_dim"])
|
||||
self.gguf_writer.add_indexer_top_k(self.hparams["index_topk"])
|
||||
|
||||
+6
-5
@@ -786,14 +786,15 @@ class Gemma4VisionAudioModel(MmprojModel):
|
||||
super().set_gguf_parameters()
|
||||
|
||||
# vision params
|
||||
assert self.hparams_vision is not None
|
||||
self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.GEMMA4V)
|
||||
self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-6))
|
||||
self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-6))
|
||||
|
||||
# audio params
|
||||
if self.hparams_audio:
|
||||
self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.GEMMA4A)
|
||||
self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["feat_in"])
|
||||
self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
|
||||
assert self.hparams_audio is not None
|
||||
self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.GEMMA4A)
|
||||
self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["feat_in"])
|
||||
self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams_audio.get("layer_norm_eps", 1e-6))
|
||||
|
||||
def is_audio_tensor(self, name: str) -> bool:
|
||||
return "audio_tower" in name or "embed_audio" in name
|
||||
|
||||
@@ -2076,6 +2076,7 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
|
||||
node_zero->src[0] = node;
|
||||
ggml_set_op_params_f32(node_zero, 0, 0.0f);
|
||||
node_zero->data = node->data;
|
||||
node_zero->buffer = node->buffer;
|
||||
node_zero->flags |= GGML_TENSOR_FLAG_COMPUTE;
|
||||
|
||||
step_cgraphs[j] = get_cgraph_aux();
|
||||
|
||||
@@ -2235,8 +2235,42 @@ static void ggml_compute_forward_fill_f32(const ggml_compute_params * params, gg
|
||||
}
|
||||
}
|
||||
|
||||
static void ggml_compute_forward_fill_f16(const ggml_compute_params * params, ggml_tensor * dst) {
|
||||
const ggml_fp16_t c = GGML_CPU_FP32_TO_FP16(ggml_get_op_params_f32(dst, 0));
|
||||
|
||||
GGML_TENSOR_LOCALS(int64_t, ne, dst, ne);
|
||||
GGML_TENSOR_LOCALS(size_t, nb, dst, nb);
|
||||
|
||||
const auto [ir0, ir1] = get_thread_range(params, dst);
|
||||
|
||||
for (int64_t ir = ir0; ir < ir1; ++ir) {
|
||||
const int64_t i03 = ir/(ne2*ne1);
|
||||
const int64_t i02 = (ir - i03*ne2*ne1)/ne1;
|
||||
const int64_t i01 = (ir - i03*ne2*ne1 - i02*ne1);
|
||||
|
||||
ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1);
|
||||
|
||||
ggml_vec_set_f16(ne0, dst_ptr, c);
|
||||
}
|
||||
}
|
||||
|
||||
void ggml_compute_forward_fill(const ggml_compute_params * params, ggml_tensor * dst) {
|
||||
ggml_compute_forward_fill_f32(params, dst);
|
||||
const ggml_tensor * src0 = dst->src[0];
|
||||
|
||||
switch (src0->type) {
|
||||
case GGML_TYPE_F32:
|
||||
{
|
||||
ggml_compute_forward_fill_f32(params, dst);
|
||||
} break;
|
||||
case GGML_TYPE_F16:
|
||||
{
|
||||
ggml_compute_forward_fill_f16(params, dst);
|
||||
} break;
|
||||
default:
|
||||
{
|
||||
GGML_ABORT("unsupported type for ggml_compute_forward_fill: %s", ggml_type_name(src0->type));
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// ggml_compute_tri
|
||||
|
||||
@@ -7,6 +7,7 @@
|
||||
#include <cstdint>
|
||||
#include <cstdlib>
|
||||
#include <memory>
|
||||
#include <mutex>
|
||||
|
||||
#if defined(GGML_USE_HIP)
|
||||
#define GGML_COMMON_DECL_HIP
|
||||
@@ -1559,6 +1560,62 @@ struct ggml_cuda_pdl_config {
|
||||
ggml_cuda_pdl_config& operator=(ggml_cuda_pdl_config&&) = delete;
|
||||
|
||||
};
|
||||
|
||||
static bool ggml_cuda_kernel_can_use_pdl(const void * kernel) {
|
||||
const int device = ggml_cuda_get_device();
|
||||
|
||||
struct cache_key {
|
||||
int device;
|
||||
const void * kernel;
|
||||
|
||||
bool operator==(const cache_key & other) const { return device == other.device && kernel == other.kernel; }
|
||||
};
|
||||
|
||||
struct cache_key_hash {
|
||||
// MurmurHash3 mixing function for better hash distribution (vs. just std::hash which in some implementations simply returns the identity)
|
||||
static size_t hash_mix(size_t x) {
|
||||
std::uint64_t y = x;
|
||||
const std::uint64_t m = 0xe9846af9b1a615d;
|
||||
|
||||
y ^= y >> 32;
|
||||
y *= m;
|
||||
y ^= y >> 32;
|
||||
y *= m;
|
||||
y ^= y >> 28;
|
||||
|
||||
return static_cast<size_t>(y);
|
||||
}
|
||||
|
||||
size_t operator()(const cache_key & key) const {
|
||||
// Use a nonzero seed to avoid mapping all-zero keys to zero
|
||||
size_t h = 42;
|
||||
h = hash_mix(h + key.device);
|
||||
h = hash_mix(h + reinterpret_cast<size_t>(key.kernel));
|
||||
return h;
|
||||
}
|
||||
};
|
||||
|
||||
static std::mutex cache_mutex;
|
||||
static std::unordered_map<cache_key, bool, cache_key_hash> cache;
|
||||
|
||||
const cache_key key = { device, kernel };
|
||||
std::lock_guard<std::mutex> lock(cache_mutex);
|
||||
const auto it = cache.find(key);
|
||||
if (it != cache.end()) {
|
||||
return it->second;
|
||||
}
|
||||
|
||||
cudaFuncAttributes attr = {};
|
||||
CUDA_CHECK(cudaFuncGetAttributes(&attr, kernel));
|
||||
|
||||
// PDL device-side primitives are emitted only for PTX versions >= 90.
|
||||
// We have to guard on a loaded kernel's PTX version so a kernel forward-JIT'ed
|
||||
// from pre-Hopper PTX to a Hopper-or-newer GPU does not opt into PDL.
|
||||
const bool can_use_pdl = attr.ptxVersion >= 90;
|
||||
cache.emplace(key, can_use_pdl);
|
||||
return can_use_pdl;
|
||||
}
|
||||
|
||||
#endif //defined(GGML_CUDA_USE_PDL)
|
||||
|
||||
|
||||
@@ -1571,8 +1628,7 @@ static __inline__ void ggml_cuda_kernel_launch(Kernel kernel, const ggml_cuda_ke
|
||||
return env == nullptr || std::atoi(env) != 0;
|
||||
}();
|
||||
|
||||
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
|
||||
if (env_pdl_enabled && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_HOPPER) {
|
||||
if (env_pdl_enabled && ggml_cuda_kernel_can_use_pdl(reinterpret_cast<const void *>(kernel))) {
|
||||
auto pdl_cfg = ggml_cuda_pdl_config(launch_params);
|
||||
|
||||
CUDA_CHECK(cudaLaunchKernelEx(&pdl_cfg.cfg, kernel, std::forward<Args>(args)... ));
|
||||
|
||||
@@ -1153,8 +1153,8 @@ void launch_fattn(
|
||||
|
||||
GGML_ASSERT(block_dim.x % warp_size == 0);
|
||||
|
||||
const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks_num, block_dim, nbytes_shared, main_stream);
|
||||
ggml_cuda_kernel_launch(fattn_kernel, launch_params,
|
||||
// disabled PDL enrollment for now due to a compiler bug.
|
||||
fattn_kernel<<<blocks_num, block_dim, nbytes_shared, main_stream>>>(
|
||||
(const char *) Q->data,
|
||||
K_data,
|
||||
V_data,
|
||||
|
||||
@@ -0,0 +1,241 @@
|
||||
#ifndef HTP_OPNODE_H
|
||||
#define HTP_OPNODE_H
|
||||
|
||||
#define GGML_COMMON_IMPL_CPP
|
||||
#include "ggml-backend-impl.h"
|
||||
#include "ggml-common.h"
|
||||
|
||||
#include <string>
|
||||
#include <vector>
|
||||
#include <stdio.h>
|
||||
#include "htp-ops.h"
|
||||
|
||||
struct htp_opnode {
|
||||
ggml_tensor * node = nullptr;
|
||||
|
||||
std::vector<ggml_tensor *> fused;
|
||||
|
||||
htp_op_code opcode = HTP_OP_INVALID;
|
||||
|
||||
ggml_op op() const {
|
||||
return node->op;
|
||||
}
|
||||
|
||||
const ggml_tensor * dst() const {
|
||||
return fused.empty() ? node : fused.back();
|
||||
}
|
||||
|
||||
const ggml_tensor * src0() const {
|
||||
return node->src[0];
|
||||
}
|
||||
|
||||
const ggml_tensor * src1() const {
|
||||
return node->src[1];
|
||||
}
|
||||
|
||||
bool is_empty() const {
|
||||
return ggml_op_is_empty(node->op);
|
||||
}
|
||||
|
||||
void add_fused(ggml_tensor * t) {
|
||||
fused.push_back(t);
|
||||
}
|
||||
|
||||
bool stackable() const {
|
||||
switch (this->op()) {
|
||||
case GGML_OP_MUL_MAT:
|
||||
case GGML_OP_MUL_MAT_ID:
|
||||
return ggml_is_quantized(this->src0()->type);
|
||||
default:
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
bool same_input(const htp_opnode& n) const {
|
||||
return n.src1() == this->src1();
|
||||
}
|
||||
|
||||
std::vector<const ggml_tensor *> get_inputs() const {
|
||||
std::vector<const ggml_tensor *> inputs;
|
||||
std::vector<const ggml_tensor *> outputs;
|
||||
outputs.push_back(node);
|
||||
for (const auto * f : fused) {
|
||||
outputs.push_back(f);
|
||||
}
|
||||
|
||||
auto contains = [&](const std::vector<const ggml_tensor *> & vec, const ggml_tensor * t) {
|
||||
for (const auto * x : vec) {
|
||||
if (x == t) return true;
|
||||
}
|
||||
return false;
|
||||
};
|
||||
|
||||
auto add_input = [&](const ggml_tensor * t) {
|
||||
if (t && !contains(outputs, t) && !contains(inputs, t)) {
|
||||
inputs.push_back(t);
|
||||
}
|
||||
};
|
||||
|
||||
for (int i = 0; i < GGML_MAX_SRC && node->src[i]; i++) {
|
||||
add_input(node->src[i]);
|
||||
}
|
||||
for (const auto * f : fused) {
|
||||
for (int i = 0; i < GGML_MAX_SRC && f->src[i]; i++) {
|
||||
add_input(f->src[i]);
|
||||
}
|
||||
}
|
||||
return inputs;
|
||||
}
|
||||
|
||||
std::string op_name() const {
|
||||
if (fused.empty()) {
|
||||
return ggml_op_desc(node);
|
||||
}
|
||||
std::string name = ggml_op_desc(node);
|
||||
for (const auto * f : fused) {
|
||||
name += "+";
|
||||
name += ggml_op_desc(f);
|
||||
}
|
||||
return name;
|
||||
}
|
||||
};
|
||||
|
||||
struct htp_opformat {
|
||||
char strides[64 * GGML_MAX_SRC];
|
||||
char dims[64 * GGML_MAX_SRC];
|
||||
char types[16 * GGML_MAX_SRC];
|
||||
char buffs[64 * GGML_MAX_SRC];
|
||||
char names[64 * GGML_MAX_SRC];
|
||||
|
||||
int format_tensor_dims(char * str, const struct ggml_tensor * t) {
|
||||
if (t->ne[2] == 1 && t->ne[3] == 1) {
|
||||
return sprintf(str, "%d:%d", (int) t->ne[0], (int) t->ne[1]);
|
||||
} else {
|
||||
return sprintf(str, "%d:%d:%d:%d", (int) t->ne[0], (int) t->ne[1], (int) t->ne[2], (int) t->ne[3]);
|
||||
}
|
||||
}
|
||||
|
||||
void format_op_dims(char * str, const htp_opnode & node) {
|
||||
char * p = str;
|
||||
auto inputs = node.get_inputs();
|
||||
|
||||
if (!inputs.empty()) {
|
||||
p += format_tensor_dims(p, inputs[0]);
|
||||
|
||||
for (size_t i = 1; i < inputs.size(); i++) {
|
||||
p += sprintf(p, " x ");
|
||||
p += format_tensor_dims(p, inputs[i]);
|
||||
}
|
||||
|
||||
p += sprintf(p, " -> ");
|
||||
}
|
||||
|
||||
char self[64];
|
||||
format_tensor_dims(self, node.dst());
|
||||
p += sprintf(p, "%s", self);
|
||||
}
|
||||
|
||||
int format_tensor_strides(char * str, const struct ggml_tensor * t) {
|
||||
const char * c = ggml_is_contiguous(t) ? "" : "!";
|
||||
|
||||
if (t->ne[2] == 1 && t->ne[3] == 1) {
|
||||
return sprintf(str, "%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], c);
|
||||
} else {
|
||||
return sprintf(str, "%zu:%zu:%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], (size_t) t->nb[2], (size_t) t->nb[3], c);
|
||||
}
|
||||
}
|
||||
|
||||
void format_op_strides(char * str, const htp_opnode & node) {
|
||||
char * p = str;
|
||||
auto inputs = node.get_inputs();
|
||||
|
||||
if (!inputs.empty()) {
|
||||
p += format_tensor_strides(p, inputs[0]);
|
||||
|
||||
for (size_t i = 1; i < inputs.size(); i++) {
|
||||
p += sprintf(p, " x ");
|
||||
p += format_tensor_strides(p, inputs[i]);
|
||||
}
|
||||
|
||||
p += sprintf(p, " -> ");
|
||||
}
|
||||
|
||||
char self[64];
|
||||
format_tensor_strides(self, node.dst());
|
||||
p += sprintf(p, "%s", self);
|
||||
}
|
||||
|
||||
void format_op_types(char * str, const htp_opnode & node) {
|
||||
char * p = str;
|
||||
auto inputs = node.get_inputs();
|
||||
|
||||
if (!inputs.empty()) {
|
||||
p += sprintf(p, "%s", ggml_type_name(inputs[0]->type));
|
||||
|
||||
for (size_t i = 1; i < inputs.size(); i++) {
|
||||
p += sprintf(p, " x ");
|
||||
p += sprintf(p, "%s", ggml_type_name(inputs[i]->type));
|
||||
}
|
||||
|
||||
p += sprintf(p, " -> ");
|
||||
}
|
||||
|
||||
p += sprintf(p, "%s", ggml_type_name(node.dst()->type));
|
||||
}
|
||||
|
||||
const char * tensor_buff_name(const struct ggml_tensor * t) {
|
||||
if (t->buffer) {
|
||||
return ggml_backend_buffer_name(t->buffer);
|
||||
}
|
||||
return "NONE";
|
||||
}
|
||||
|
||||
void format_op_buffs(char * str, const htp_opnode & node) {
|
||||
char * p = str;
|
||||
auto inputs = node.get_inputs();
|
||||
|
||||
if (!inputs.empty()) {
|
||||
p += sprintf(p, "%s", tensor_buff_name(inputs[0]));
|
||||
|
||||
for (size_t i = 1; i < inputs.size(); i++) {
|
||||
p += sprintf(p, " x ");
|
||||
p += sprintf(p, "%s", tensor_buff_name(inputs[i]));
|
||||
}
|
||||
|
||||
p += sprintf(p, " -> ");
|
||||
}
|
||||
|
||||
p += sprintf(p, "%s", tensor_buff_name(node.dst()));
|
||||
}
|
||||
|
||||
void format_op_names(char * str, const htp_opnode & node) {
|
||||
char * p = str;
|
||||
auto inputs = node.get_inputs();
|
||||
|
||||
if (!inputs.empty()) {
|
||||
p += sprintf(p, "%s", inputs[0]->name);
|
||||
|
||||
for (size_t i = 1; i < inputs.size(); i++) {
|
||||
p += sprintf(p, " x ");
|
||||
p += sprintf(p, "%s", inputs[i]->name);
|
||||
}
|
||||
|
||||
p += sprintf(p, " -> ");
|
||||
}
|
||||
|
||||
p += sprintf(p, "%s", node.dst()->name);
|
||||
}
|
||||
|
||||
void format(const htp_opnode & node) {
|
||||
format_op_dims(dims, node);
|
||||
format_op_strides(strides, node);
|
||||
format_op_types(types, node);
|
||||
format_op_buffs(buffs, node);
|
||||
format_op_names(names, node);
|
||||
}
|
||||
|
||||
htp_opformat() {}
|
||||
htp_opformat(const htp_opnode & node) { format(node); }
|
||||
};
|
||||
|
||||
#endif // HTP_OPNODE_H
|
||||
+1
-1
@@ -5239,7 +5239,7 @@ static struct ggml_tensor * ggml_fill_impl(
|
||||
struct ggml_tensor * a,
|
||||
float c,
|
||||
bool inplace) {
|
||||
GGML_ASSERT(a->type == GGML_TYPE_F32);
|
||||
GGML_ASSERT(a->type == GGML_TYPE_F32 || a->type == GGML_TYPE_F16);
|
||||
GGML_ASSERT(ggml_is_contiguous(a));
|
||||
|
||||
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
|
||||
|
||||
@@ -451,6 +451,7 @@ class MODEL_ARCH(IntEnum):
|
||||
DEEPSEEK = auto()
|
||||
DEEPSEEK2 = auto()
|
||||
DEEPSEEK2OCR = auto()
|
||||
DEEPSEEK32 = auto()
|
||||
CHATGLM = auto()
|
||||
GLM4 = auto()
|
||||
GLM4_MOE = auto()
|
||||
@@ -811,6 +812,8 @@ class MODEL_TENSOR(IntEnum):
|
||||
V_SAM_NET_3 = auto() # Deepseek-OCR
|
||||
V_ENC_EMBD_IMGNL = auto() # Deepseek-OCR
|
||||
V_ENC_EMBD_VSEP = auto() # Deepseek-OCR
|
||||
V_RESMPL_QUERY_768 = auto() # Deepseek-OCR-2
|
||||
V_RESMPL_QUERY_1024 = auto() # Deepseek-OCR-2
|
||||
|
||||
# audio (mtmd)
|
||||
A_ENC_EMBD_POS = auto()
|
||||
@@ -967,6 +970,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
|
||||
MODEL_ARCH.DEEPSEEK: "deepseek",
|
||||
MODEL_ARCH.DEEPSEEK2: "deepseek2",
|
||||
MODEL_ARCH.DEEPSEEK2OCR: "deepseek2-ocr",
|
||||
MODEL_ARCH.DEEPSEEK32: "deepseek32",
|
||||
MODEL_ARCH.CHATGLM: "chatglm",
|
||||
MODEL_ARCH.GLM4: "glm4",
|
||||
MODEL_ARCH.GLM4_MOE: "glm4moe",
|
||||
@@ -1327,6 +1331,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
|
||||
MODEL_TENSOR.V_SAM_NET_3: "v.sam.net_3",
|
||||
MODEL_TENSOR.V_ENC_EMBD_IMGNL: "v.image_newline", # Deepseek-OCR
|
||||
MODEL_TENSOR.V_ENC_EMBD_VSEP: "v.view_seperator", # Deepseek-OCR
|
||||
MODEL_TENSOR.V_RESMPL_QUERY_768: "v.resample_query_768", # Deepseek-OCR-2 qwen2
|
||||
MODEL_TENSOR.V_RESMPL_QUERY_1024: "v.resample_query_1024", # Deepseek-OCR-2 qwen2
|
||||
# audio (mtmd)
|
||||
# note: all audio tensor names must use prefix "a." or "mm.a."
|
||||
MODEL_TENSOR.A_ENC_EMBD_POS: "a.position_embd",
|
||||
@@ -1505,6 +1511,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
|
||||
MODEL_TENSOR.V_SAM_NECK,
|
||||
MODEL_TENSOR.V_SAM_NET_2,
|
||||
MODEL_TENSOR.V_SAM_NET_3,
|
||||
MODEL_TENSOR.V_RESMPL_QUERY_768,
|
||||
MODEL_TENSOR.V_RESMPL_QUERY_1024,
|
||||
# audio
|
||||
MODEL_TENSOR.A_ENC_EMBD_POS,
|
||||
MODEL_TENSOR.A_ENC_EMBD_NORM,
|
||||
@@ -2930,6 +2938,46 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
|
||||
MODEL_TENSOR.FFN_UP_SHEXP,
|
||||
MODEL_TENSOR.FFN_EXP_PROBS_B,
|
||||
],
|
||||
MODEL_ARCH.DEEPSEEK32: [
|
||||
MODEL_TENSOR.TOKEN_EMBD,
|
||||
MODEL_TENSOR.OUTPUT_NORM,
|
||||
MODEL_TENSOR.OUTPUT,
|
||||
MODEL_TENSOR.ROPE_FREQS,
|
||||
MODEL_TENSOR.ATTN_NORM,
|
||||
MODEL_TENSOR.ATTN_Q,
|
||||
MODEL_TENSOR.ATTN_Q_A,
|
||||
MODEL_TENSOR.ATTN_Q_B,
|
||||
MODEL_TENSOR.ATTN_KV_A_MQA,
|
||||
MODEL_TENSOR.ATTN_K_B,
|
||||
MODEL_TENSOR.ATTN_V_B,
|
||||
MODEL_TENSOR.ATTN_Q_A_NORM,
|
||||
MODEL_TENSOR.ATTN_KV_A_NORM,
|
||||
MODEL_TENSOR.ATTN_OUT,
|
||||
MODEL_TENSOR.ATTN_ROT_EMBD,
|
||||
MODEL_TENSOR.FFN_GATE_INP,
|
||||
MODEL_TENSOR.FFN_NORM,
|
||||
MODEL_TENSOR.FFN_GATE,
|
||||
MODEL_TENSOR.FFN_DOWN,
|
||||
MODEL_TENSOR.FFN_UP,
|
||||
MODEL_TENSOR.FFN_GATE_EXP,
|
||||
MODEL_TENSOR.FFN_DOWN_EXP,
|
||||
MODEL_TENSOR.FFN_UP_EXP,
|
||||
MODEL_TENSOR.FFN_GATE_SHEXP,
|
||||
MODEL_TENSOR.FFN_DOWN_SHEXP,
|
||||
MODEL_TENSOR.FFN_UP_SHEXP,
|
||||
MODEL_TENSOR.FFN_EXP_PROBS_B,
|
||||
MODEL_TENSOR.INDEXER_K_NORM,
|
||||
MODEL_TENSOR.INDEXER_PROJ,
|
||||
MODEL_TENSOR.INDEXER_ATTN_K,
|
||||
MODEL_TENSOR.INDEXER_ATTN_Q_B,
|
||||
# NextN/MTP tensors - preserved but unused
|
||||
MODEL_TENSOR.NEXTN_EH_PROJ,
|
||||
MODEL_TENSOR.NEXTN_EMBED_TOKENS,
|
||||
MODEL_TENSOR.NEXTN_ENORM,
|
||||
MODEL_TENSOR.NEXTN_HNORM,
|
||||
MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
|
||||
MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
|
||||
],
|
||||
MODEL_ARCH.ERNIE4_5_MOE: [
|
||||
MODEL_TENSOR.TOKEN_EMBD,
|
||||
MODEL_TENSOR.OUTPUT_NORM,
|
||||
@@ -4077,6 +4125,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
|
||||
MODEL_TENSOR.ROPE_FREQS,
|
||||
MODEL_TENSOR.ATTN_ROT_EMBD,
|
||||
],
|
||||
MODEL_ARCH.DEEPSEEK32: [
|
||||
MODEL_TENSOR.ROPE_FREQS,
|
||||
MODEL_TENSOR.ATTN_ROT_EMBD,
|
||||
],
|
||||
MODEL_ARCH.CHATGLM: [
|
||||
MODEL_TENSOR.ROPE_FREQS,
|
||||
],
|
||||
@@ -4283,6 +4335,7 @@ class VisionProjectorType:
|
||||
JANUS_PRO = "janus_pro"
|
||||
DOTSOCR = "dots_ocr"
|
||||
DEEPSEEKOCR = "deepseekocr"
|
||||
DEEPSEEKOCR2 = "deepseekocr2"
|
||||
LFM2A = "lfm2a" # audio
|
||||
MUSIC_FLAMINGO = "musicflamingo" # audio
|
||||
GLM4V = "glm4v"
|
||||
|
||||
@@ -1485,6 +1485,7 @@ class TensorNameMap:
|
||||
"siglip2.vision_model.encoder.layers.{bid}.self_attn.q_proj", # youtuvl
|
||||
"model.vision_model.transformer.layers.{bid}.self_attn.q_proj", # Deepseek-OCR CLIP, generated
|
||||
"vision_model.model.layers.{bid}.self_attn.q_proj.linear", # gemma4
|
||||
"model.qwen2_model.model.model.layers.{bid}.self_attn.q_proj" # Deepseek-OCR-2 qwen2
|
||||
),
|
||||
|
||||
MODEL_TENSOR.V_ENC_ATTN_Q_NORM: (
|
||||
@@ -1509,6 +1510,7 @@ class TensorNameMap:
|
||||
"model.vision_model.transformer.layers.{bid}.self_attn.k_proj", # Deepseek-OCR CLIP, generated
|
||||
"siglip2.vision_model.encoder.layers.{bid}.self_attn.k_proj",
|
||||
"vision_model.model.layers.{bid}.self_attn.k_proj.linear", # gemma4
|
||||
"model.qwen2_model.model.model.layers.{bid}.self_attn.k_proj" # Deepseek-OCR-2 qwen2
|
||||
),
|
||||
|
||||
MODEL_TENSOR.V_ENC_ATTN_K_NORM: (
|
||||
@@ -1533,6 +1535,7 @@ class TensorNameMap:
|
||||
"siglip2.vision_model.encoder.layers.{bid}.self_attn.v_proj",
|
||||
"model.vision_model.transformer.layers.{bid}.self_attn.v_proj", # Deepseek-OCR CLIP, generated
|
||||
"vision_model.model.layers.{bid}.self_attn.v_proj.linear", # gemma4
|
||||
"model.qwen2_model.model.model.layers.{bid}.self_attn.v_proj" # Deepseek-OCR-2 qwen2
|
||||
),
|
||||
|
||||
MODEL_TENSOR.V_ENC_INPUT_NORM: (
|
||||
@@ -1554,6 +1557,7 @@ class TensorNameMap:
|
||||
"vision_model.radio_model.model.blocks.{bid}.norm1", # Nemotron Nano v2 VL
|
||||
"vision_tower.blocks.{bid}.norm1", # dots.ocr
|
||||
"vision_model.transformer.resblocks.{bid}.ln_1", # Step3-VL
|
||||
"model.qwen2_model.model.model.layers.{bid}.input_layernorm", # Deepseek-OCR-2 qwen2
|
||||
),
|
||||
|
||||
MODEL_TENSOR.V_ENC_ATTN_O: (
|
||||
@@ -1574,6 +1578,7 @@ class TensorNameMap:
|
||||
"model.vision_model.transformer.layers.{bid}.self_attn.out_proj", # Deepseek-OCR CLIP
|
||||
"siglip2.vision_model.encoder.layers.{bid}.self_attn.out_proj", # youtuvl
|
||||
"vision_model.radio_model.model.blocks.{bid}.attn.proj", # Nemotron Nano v2 VL
|
||||
"model.qwen2_model.model.model.layers.{bid}.self_attn.o_proj", # Deepseek-OCR-2 qwen2
|
||||
"vision_model.model.layers.{bid}.self_attn.o_proj.linear", # gemma4
|
||||
"vision_tower.blocks.{bid}.attn.proj", # dots.ocr
|
||||
"vision_model.transformer.resblocks.{bid}.attn.out_proj", # Step3-VL
|
||||
@@ -1603,6 +1608,7 @@ class TensorNameMap:
|
||||
"vision_model.model.layers.{bid}.pre_feedforward_layernorm", # gemma4
|
||||
"vision_tower.blocks.{bid}.norm2", # dots.ocr
|
||||
"vision_model.transformer.resblocks.{bid}.ln_2", # Step3-VL
|
||||
"model.qwen2_model.model.model.layers.{bid}.post_attention_layernorm", # Deepseek-OCR-2 qwen2
|
||||
),
|
||||
|
||||
MODEL_TENSOR.V_ENC_FFN_UP: (
|
||||
@@ -1625,6 +1631,7 @@ class TensorNameMap:
|
||||
"vision_model.radio_model.model.blocks.{bid}.mlp.fc1", # Nemotron Nano v2 VL
|
||||
"vision_model.model.layers.{bid}.mlp.up_proj", # gemma4
|
||||
"vision_model.transformer.resblocks.{bid}.mlp.c_fc", # Step3-VL
|
||||
"model.qwen2_model.model.model.layers.{bid}.mlp.up_proj", # Deepseek-OCR-2 qwen2
|
||||
),
|
||||
|
||||
MODEL_TENSOR.V_ENC_FFN_GATE: (
|
||||
@@ -1632,6 +1639,7 @@ class TensorNameMap:
|
||||
"vision_encoder.transformer.layers.{bid}.feed_forward.w1", # pixtral
|
||||
"visual.blocks.{bid}.mlp.gate_proj", # qwen2.5vl
|
||||
"vision_model.model.layers.{bid}.mlp.gate_proj", # gemma4
|
||||
"model.qwen2_model.model.model.layers.{bid}.mlp.gate_proj", # Deepseek-OCR-2 qwen2
|
||||
),
|
||||
|
||||
MODEL_TENSOR.V_ENC_FFN_DOWN: (
|
||||
@@ -1652,6 +1660,7 @@ class TensorNameMap:
|
||||
"model.vision_model.transformer.layers.{bid}.mlp.fc2", # Deepseek-OCR CLIP
|
||||
"siglip2.vision_model.encoder.layers.{bid}.mlp.fc2",
|
||||
"vision_model.radio_model.model.blocks.{bid}.mlp.fc2", # Nemotron Nano v2 VL
|
||||
"model.qwen2_model.model.model.layers.{bid}.mlp.down_proj" , # Deepseek-OCR-2 qwen2
|
||||
"vision_model.model.layers.{bid}.mlp.down_proj", # gemma4
|
||||
"vision_model.transformer.resblocks.{bid}.mlp.c_proj", # Step3-VL
|
||||
),
|
||||
@@ -1699,6 +1708,7 @@ class TensorNameMap:
|
||||
"vision_tower.encoder.final_layernorm", # kimi-vl
|
||||
"visual.post_layernorm", # glm4v
|
||||
"siglip2.vision_model.post_layernorm",
|
||||
"model.qwen2_model.model.model.norm", # Deepseek-OCR-2 qwen2
|
||||
),
|
||||
|
||||
MODEL_TENSOR.V_MM_POST_NORM: (
|
||||
@@ -1879,6 +1889,14 @@ class TensorNameMap:
|
||||
"model.sam_model.net_3",
|
||||
),
|
||||
|
||||
MODEL_TENSOR.V_RESMPL_QUERY_768: (
|
||||
"model.qwen2_model.query_768", # Deepseek-OCR-2 qwen2
|
||||
),
|
||||
|
||||
MODEL_TENSOR.V_RESMPL_QUERY_1024: (
|
||||
"model.qwen2_model.query_1024", # Deepseek-OCR-2 qwen2
|
||||
),
|
||||
|
||||
MODEL_TENSOR.V_MM_POST_FC_NORM: (
|
||||
"model.vision.linear_proj.norm1", # cogvlm
|
||||
),
|
||||
|
||||
@@ -75,6 +75,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
|
||||
{ LLM_ARCH_DEEPSEEK, "deepseek" },
|
||||
{ LLM_ARCH_DEEPSEEK2, "deepseek2" },
|
||||
{ LLM_ARCH_DEEPSEEK2OCR, "deepseek2-ocr" },
|
||||
{ LLM_ARCH_DEEPSEEK32, "deepseek32" },
|
||||
{ LLM_ARCH_CHATGLM, "chatglm" },
|
||||
{ LLM_ARCH_GLM4, "glm4" },
|
||||
{ LLM_ARCH_GLM4_MOE, "glm4moe" },
|
||||
@@ -904,6 +905,7 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
|
||||
case LLM_ARCH_OLMO2:
|
||||
case LLM_ARCH_OLMOE:
|
||||
case LLM_ARCH_DEEPSEEK2:
|
||||
case LLM_ARCH_DEEPSEEK32:
|
||||
case LLM_ARCH_GLM_DSA:
|
||||
case LLM_ARCH_BITNET:
|
||||
case LLM_ARCH_T5:
|
||||
|
||||
@@ -79,6 +79,7 @@ enum llm_arch {
|
||||
LLM_ARCH_DEEPSEEK,
|
||||
LLM_ARCH_DEEPSEEK2,
|
||||
LLM_ARCH_DEEPSEEK2OCR,
|
||||
LLM_ARCH_DEEPSEEK32,
|
||||
LLM_ARCH_CHATGLM,
|
||||
LLM_ARCH_GLM4,
|
||||
LLM_ARCH_GLM4_MOE,
|
||||
|
||||
+225
-51
@@ -7,6 +7,7 @@
|
||||
|
||||
#include "llama-kv-cache.h"
|
||||
#include "llama-kv-cache-iswa.h"
|
||||
#include "llama-kv-cache-dsa.h"
|
||||
#include "llama-memory-hybrid.h"
|
||||
#include "llama-memory-hybrid-iswa.h"
|
||||
#include "llama-memory-recurrent.h"
|
||||
@@ -29,7 +30,10 @@ static ggml_tensor * build_attn_inp_kq_mask(
|
||||
const auto n_tokens = ubatch.n_tokens;
|
||||
const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
|
||||
|
||||
ggml_tensor * res = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
|
||||
// flash attention requires an f16 mask
|
||||
const auto type = cparams.flash_attn ? GGML_TYPE_F16 : GGML_TYPE_F32;
|
||||
|
||||
ggml_tensor * res = ggml_new_tensor_4d(ctx, type, n_kv, n_tokens/n_stream, 1, n_stream);
|
||||
ggml_set_input(res);
|
||||
ggml_set_name(res, "attn_inp_kq_mask");
|
||||
|
||||
@@ -102,6 +106,39 @@ bool llm_graph_input_embd::can_reuse(const llm_graph_params & params) {
|
||||
return res;
|
||||
}
|
||||
|
||||
void llm_graph_input_embd_h::set_input(const llama_ubatch * ubatch) {
|
||||
const int64_t n_tokens = ubatch->n_tokens;
|
||||
|
||||
if (ubatch->token) {
|
||||
ggml_backend_tensor_set(tokens, ubatch->token, 0, n_tokens*ggml_element_size(tokens));
|
||||
} else {
|
||||
// note: mtmd embedding input goes through here
|
||||
GGML_ASSERT(ubatch->embd);
|
||||
GGML_ASSERT(n_embd == embd->ne[0]);
|
||||
|
||||
ggml_backend_tensor_set(embd, ubatch->embd, 0, n_tokens*n_embd*ggml_element_size(h));
|
||||
}
|
||||
|
||||
// TODO: extend llama_ubatch to differentiate between token embeddings and hidden states
|
||||
// for now, we assume that the hidden state is always provided as an embedding
|
||||
// ref: https://github.com/ggml-org/llama.cpp/pull/23643
|
||||
if (ubatch->embd) {
|
||||
GGML_ASSERT(n_embd == h->ne[0]);
|
||||
|
||||
ggml_backend_tensor_set(h, ubatch->embd, 0, n_tokens*n_embd*ggml_element_size(h));
|
||||
}
|
||||
}
|
||||
|
||||
bool llm_graph_input_embd_h::can_reuse(const llm_graph_params & params) {
|
||||
bool res = true;
|
||||
|
||||
res &= (!params.ubatch.token) || (tokens && tokens->ne[0] == params.ubatch.n_tokens);
|
||||
res &= (!params.ubatch.embd) || (embd && embd->ne[1] == params.ubatch.n_tokens);
|
||||
res &= (!params.ubatch.embd) || (h && h->ne[1] == params.ubatch.n_tokens);
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
|
||||
if (ubatch->pos && pos) {
|
||||
const int64_t n_tokens = ubatch->n_tokens;
|
||||
@@ -348,7 +385,8 @@ void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
|
||||
}
|
||||
}
|
||||
|
||||
static void print_mask(const float * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
|
||||
template <typename T>
|
||||
static void print_mask(const T * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
|
||||
LLAMA_LOG_DEBUG("%s: === Attention mask ===\n", __func__);
|
||||
const char * swa_type_str = "unknown";
|
||||
|
||||
@@ -372,7 +410,7 @@ static void print_mask(const float * data, int64_t n_tokens, int64_t n_kv, int64
|
||||
for (int i = 0; i < std::min((int64_t)20, n_tokens); ++i) {
|
||||
LLAMA_LOG_DEBUG(" %2d ", i);
|
||||
for (int j = 0; j < std::min((int64_t)20, n_kv); ++j) {
|
||||
float val = data[i * n_kv + j];
|
||||
float val = llama_cast<float>(data[i * n_kv + j]);
|
||||
if (val == -INFINITY) {
|
||||
LLAMA_LOG_DEBUG(" ∞");
|
||||
} else {
|
||||
@@ -387,7 +425,10 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
|
||||
const int64_t n_kv = ubatch->n_tokens;
|
||||
const int64_t n_tokens = ubatch->n_tokens;
|
||||
|
||||
const auto fill_mask = [&](float * data, int n_swa, llama_swa_type swa_type) {
|
||||
const auto fill_mask = [&](auto * data, int64_t ne, int n_swa, llama_swa_type swa_type) {
|
||||
using T = std::remove_reference_t<decltype(*data)>;
|
||||
std::fill(data, data + ne, llama_cast<T>(-INFINITY));
|
||||
|
||||
for (int i1 = 0; i1 < n_tokens; ++i1) {
|
||||
const llama_seq_id s1 = ubatch->seq_id[i1][0];
|
||||
const llama_pos p1 = ubatch->pos[i1];
|
||||
@@ -413,38 +454,30 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
|
||||
continue;
|
||||
}
|
||||
|
||||
data[idst + i0] = hparams.use_alibi ? -std::abs(p0 - p1) : 0.0f;
|
||||
data[idst + i0] = llama_cast<T>(hparams.use_alibi ? -std::abs(p0 - p1) : 0.0f);
|
||||
}
|
||||
}
|
||||
|
||||
if (debug) {
|
||||
print_mask(data, n_tokens, n_kv, n_swa, swa_type);
|
||||
}
|
||||
};
|
||||
|
||||
{
|
||||
GGML_ASSERT(self_kq_mask);
|
||||
GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer));
|
||||
|
||||
float * data = (float *) self_kq_mask->data;
|
||||
|
||||
std::fill(data, data + ggml_nelements(self_kq_mask), -INFINITY);
|
||||
|
||||
fill_mask(data, 0, LLAMA_SWA_TYPE_NONE);
|
||||
|
||||
if (debug) {
|
||||
print_mask(data, n_tokens, n_kv, 0, LLAMA_SWA_TYPE_NONE);
|
||||
}
|
||||
GGML_ASSERT(self_kq_mask);
|
||||
GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer));
|
||||
if (self_kq_mask->type == GGML_TYPE_F16) {
|
||||
fill_mask((ggml_fp16_t *) self_kq_mask->data, ggml_nelements(self_kq_mask), 0, LLAMA_SWA_TYPE_NONE);
|
||||
} else {
|
||||
fill_mask((float *) self_kq_mask->data, ggml_nelements(self_kq_mask), 0, LLAMA_SWA_TYPE_NONE);
|
||||
}
|
||||
|
||||
if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
|
||||
GGML_ASSERT(self_kq_mask_swa);
|
||||
GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask_swa->buffer));
|
||||
|
||||
float * data = (float *) self_kq_mask_swa->data;
|
||||
|
||||
std::fill(data, data + ggml_nelements(self_kq_mask_swa), -INFINITY);
|
||||
|
||||
fill_mask(data, hparams.n_swa, hparams.swa_type);
|
||||
|
||||
if (debug) {
|
||||
print_mask(data, n_tokens, n_kv, hparams.n_swa, hparams.swa_type);
|
||||
if (self_kq_mask_swa->type == GGML_TYPE_F16) {
|
||||
fill_mask((ggml_fp16_t *) self_kq_mask_swa->data, ggml_nelements(self_kq_mask_swa), hparams.n_swa, hparams.swa_type);
|
||||
} else {
|
||||
fill_mask((float *) self_kq_mask_swa->data, ggml_nelements(self_kq_mask_swa), hparams.n_swa, hparams.swa_type);
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -499,6 +532,34 @@ bool llm_graph_input_attn_k::can_reuse(const llm_graph_params & params) {
|
||||
return res;
|
||||
}
|
||||
|
||||
void llm_graph_input_attn_k_dsa::set_input(const llama_ubatch * ubatch) {
|
||||
mctx->get_mla()->set_input_k_idxs(self_k_idxs_mla, ubatch);
|
||||
|
||||
mctx->get_mla()->set_input_kq_mask(self_kq_mask_mla, ubatch, cparams.causal_attn);
|
||||
|
||||
mctx->get_lid()->set_input_k_idxs(self_k_idxs_lid, ubatch);
|
||||
|
||||
mctx->get_lid()->set_input_kq_mask(self_kq_mask_lid, ubatch, cparams.causal_attn);
|
||||
|
||||
mctx->get_lid()->set_input_k_rot(self_k_rot_lid);
|
||||
}
|
||||
|
||||
bool llm_graph_input_attn_k_dsa::can_reuse(const llm_graph_params & params) {
|
||||
const auto * mctx = static_cast<const llama_kv_cache_dsa_context *>(params.mctx);
|
||||
|
||||
this->mctx = mctx;
|
||||
|
||||
bool res = true;
|
||||
|
||||
res &= self_k_idxs_mla->ne[0] == params.ubatch.n_tokens;
|
||||
res &= self_k_idxs_lid->ne[0] == params.ubatch.n_tokens;
|
||||
|
||||
res &= can_reuse_kq_mask(self_kq_mask_mla, mctx->get_mla(), params.ubatch, params.cparams);
|
||||
res &= can_reuse_kq_mask(self_kq_mask_lid, mctx->get_lid(), params.ubatch, params.cparams);
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
void llm_graph_input_attn_kv_iswa::set_input(const llama_ubatch * ubatch) {
|
||||
// base tensors may not be allocated if there are no non-SWA attention layers
|
||||
if (self_k_idxs && self_k_idxs->buffer) {
|
||||
@@ -568,23 +629,30 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
|
||||
GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer));
|
||||
GGML_ASSERT(!ubatch->equal_seqs()); // TODO: use ubatch->n_seqs instead of failing
|
||||
|
||||
float * data = (float *) cross_kq_mask->data;
|
||||
const auto fill_mask = [&](auto * data) {
|
||||
using T = std::remove_reference_t<decltype(*data)>;
|
||||
for (int i = 0; i < n_tokens; ++i) {
|
||||
GGML_ASSERT(!cross->seq_ids_enc.empty() && "llama_encode must be called first");
|
||||
for (int j = 0; j < n_enc; ++j) {
|
||||
float f = -INFINITY;
|
||||
|
||||
for (int i = 0; i < n_tokens; ++i) {
|
||||
GGML_ASSERT(!cross->seq_ids_enc.empty() && "llama_encode must be called first");
|
||||
for (int j = 0; j < n_enc; ++j) {
|
||||
float f = -INFINITY;
|
||||
for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
|
||||
const llama_seq_id seq_id = ubatch->seq_id[i][s];
|
||||
|
||||
for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
|
||||
const llama_seq_id seq_id = ubatch->seq_id[i][s];
|
||||
|
||||
if (cross->seq_ids_enc[j].find(seq_id) != cross->seq_ids_enc[j].end()) {
|
||||
f = 0.0f;
|
||||
if (cross->seq_ids_enc[j].find(seq_id) != cross->seq_ids_enc[j].end()) {
|
||||
f = 0.0f;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
data[i*n_enc + j] = f;
|
||||
data[i*n_enc + j] = llama_cast<T>(f);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
if (cross_kq_mask->type == GGML_TYPE_F16) {
|
||||
fill_mask((ggml_fp16_t *) cross_kq_mask->data);
|
||||
} else {
|
||||
fill_mask((float *) cross_kq_mask->data);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -2088,17 +2156,20 @@ ggml_tensor * llm_graph_context::build_attn_mha(
|
||||
llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() const {
|
||||
auto inp = std::make_unique<llm_graph_input_attn_no_cache>(hparams, cparams);
|
||||
|
||||
// flash attention requires an f16 mask
|
||||
const auto type_mask = cparams.flash_attn ? GGML_TYPE_F16 : GGML_TYPE_F32;
|
||||
|
||||
// note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch
|
||||
inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens, 1, 1);
|
||||
inp->self_kq_mask = ggml_new_tensor_4d(ctx0, type_mask, n_tokens, n_tokens, 1, 1);
|
||||
ggml_set_input(inp->self_kq_mask);
|
||||
|
||||
inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
|
||||
inp->self_kq_mask_cnv = inp->self_kq_mask;
|
||||
|
||||
if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
|
||||
inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens, 1, 1);
|
||||
inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, type_mask, n_tokens, n_tokens, 1, 1);
|
||||
ggml_set_input(inp->self_kq_mask_swa);
|
||||
|
||||
inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
|
||||
inp->self_kq_mask_swa_cnv = inp->self_kq_mask_swa;
|
||||
} else {
|
||||
inp->self_kq_mask_swa = nullptr;
|
||||
inp->self_kq_mask_swa_cnv = nullptr;
|
||||
@@ -2175,7 +2246,7 @@ static std::unique_ptr<llm_graph_input_attn_kv> build_attn_inp_kv_impl(
|
||||
inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch);
|
||||
|
||||
inp->self_kq_mask = build_attn_inp_kq_mask(ctx0, mctx_cur, ubatch, cparams);
|
||||
inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
|
||||
inp->self_kq_mask_cnv = inp->self_kq_mask;
|
||||
}
|
||||
|
||||
inp->self_k_rot = mctx_cur->build_input_k_rot(ctx0);
|
||||
@@ -2282,7 +2353,7 @@ static std::unique_ptr<llm_graph_input_attn_k> build_attn_inp_k_impl(
|
||||
inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
|
||||
|
||||
inp->self_kq_mask = build_attn_inp_kq_mask(ctx0, mctx_cur, ubatch, cparams);
|
||||
inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
|
||||
inp->self_kq_mask_cnv = inp->self_kq_mask;
|
||||
}
|
||||
|
||||
return inp;
|
||||
@@ -2354,6 +2425,82 @@ ggml_tensor * llm_graph_context::build_attn(
|
||||
return cur;
|
||||
}
|
||||
|
||||
ggml_tensor * llm_graph_context::build_attn(
|
||||
llm_graph_input_attn_k_dsa * inp,
|
||||
ggml_tensor * wo,
|
||||
ggml_tensor * wo_b,
|
||||
ggml_tensor * wo_s,
|
||||
ggml_tensor * q_cur,
|
||||
ggml_tensor * k_cur,
|
||||
ggml_tensor * v_cur,
|
||||
ggml_tensor * kq_b,
|
||||
ggml_tensor * sinks,
|
||||
ggml_tensor * v_mla,
|
||||
ggml_tensor * top_k,
|
||||
float kq_scale,
|
||||
int il) const {
|
||||
// these nodes are added to the graph together so that they are not reordered
|
||||
// by doing so, the number of splits in the graph is reduced
|
||||
// expand k later to enable rope fusion which directly writes into k-v cache
|
||||
ggml_build_forward_expand(gf, q_cur);
|
||||
ggml_build_forward_expand(gf, v_cur);
|
||||
ggml_build_forward_expand(gf, k_cur);
|
||||
|
||||
const auto * mctx_cur = inp->mctx->get_mla();
|
||||
|
||||
// store to KV cache
|
||||
{
|
||||
const auto & k_idxs = inp->get_k_idxs_mla();
|
||||
|
||||
ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
|
||||
}
|
||||
|
||||
const auto & kq_mask = inp->get_kq_mask_mla();
|
||||
|
||||
// prepare new kq mask - starts filled with -INFINITY
|
||||
ggml_tensor * kq_mask_all = ggml_fill(ctx0, kq_mask, -INFINITY);
|
||||
|
||||
// reshape KQ mask into tensor with rows of size 1:
|
||||
// [n_kv, n_batch, 1, n_stream] -> [1, n_kv, n_batch, n_stream]
|
||||
kq_mask_all = ggml_view_4d(ctx0, kq_mask_all, 1, kq_mask_all->ne[0], kq_mask_all->ne[1], kq_mask_all->ne[3], kq_mask_all->nb[0], kq_mask_all->nb[1], kq_mask_all->nb[2], 0);
|
||||
|
||||
// reshape top_k indices: [n_top_k, n_batch, 1, n_stream] -> [n_top_k, n_batch, n_stream, 1]
|
||||
ggml_tensor * top_k_3d = ggml_view_4d(ctx0, top_k, top_k->ne[0], top_k->ne[1], top_k->ne[3], 1, top_k->nb[1], top_k->nb[2], top_k->ne[3]*top_k->nb[3], 0);
|
||||
|
||||
// prepare zero-filled tensor with rows of size 1: [1, n_top_k, n_batch, n_stream]
|
||||
// this will be our source of zero values for unmasking top k mask elements
|
||||
ggml_tensor * zeros = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, 1, top_k_3d->ne[0], top_k_3d->ne[1], top_k_3d->ne[2]);
|
||||
zeros = ggml_fill(ctx0, zeros, 0.0f);
|
||||
|
||||
// modify KQ mask by unmasking elements that are in top_k indices
|
||||
// ggml_set_rows([1, n_kv, n_batch, n_stream], [1, n_top_k, n_batch, n_stream], [n_top_k, n_batch, n_stream, 1])
|
||||
ggml_tensor * kq_mask_top_k = ggml_set_rows(ctx0, kq_mask_all, zeros, top_k_3d);
|
||||
|
||||
// reshape to restore the original shape of KQ mask:
|
||||
// [1, n_kv, n_batch, n_stream] -> [n_kv, n_batch, 1, n_stream]
|
||||
kq_mask_top_k = ggml_view_4d(ctx0, kq_mask_top_k, kq_mask_top_k->ne[1], kq_mask_top_k->ne[2], 1, kq_mask_top_k->ne[3], kq_mask_top_k->nb[2], kq_mask_top_k->nb[3], kq_mask_top_k->nb[3], 0);
|
||||
|
||||
// combine with the original kq mask
|
||||
kq_mask_top_k = ggml_add(ctx0, kq_mask_top_k, kq_mask);
|
||||
|
||||
ggml_tensor * q = q_cur;
|
||||
ggml_tensor * k = mctx_cur->get_k(ctx0, il);
|
||||
ggml_tensor * v = ggml_view_4d(ctx0, k, v_cur->ne[0], k->ne[1], k->ne[2], k->ne[3], k->nb[1], k->nb[2], k->nb[3], 0);
|
||||
|
||||
ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask_top_k, sinks, v_mla, kq_scale, il);
|
||||
cb(cur, "kqv_out", il);
|
||||
|
||||
if (wo) {
|
||||
cur = build_lora_mm(wo, cur, wo_s);
|
||||
}
|
||||
|
||||
if (wo_b) {
|
||||
cur = ggml_add(ctx0, cur, wo_b);
|
||||
}
|
||||
|
||||
return cur;
|
||||
}
|
||||
|
||||
ggml_tensor * llm_graph_context::build_attn(
|
||||
llm_graph_input_attn_kv_iswa * inp,
|
||||
ggml_tensor * wo,
|
||||
@@ -2446,10 +2593,13 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
|
||||
|
||||
const int32_t n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
|
||||
|
||||
inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_enc, n_tokens, 1, 1);
|
||||
// flash attention requires an f16 mask
|
||||
const auto type_mask = cparams.flash_attn ? GGML_TYPE_F16 : GGML_TYPE_F32;
|
||||
|
||||
inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, type_mask, n_enc, n_tokens, 1, 1);
|
||||
ggml_set_input(inp->cross_kq_mask);
|
||||
|
||||
inp->cross_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->cross_kq_mask, GGML_TYPE_F16) : inp->cross_kq_mask;
|
||||
inp->cross_kq_mask_cnv = inp->cross_kq_mask;
|
||||
|
||||
return (llm_graph_input_attn_cross *) res->add_input(std::move(inp));
|
||||
}
|
||||
@@ -2497,6 +2647,30 @@ ggml_tensor * llm_graph_context::build_attn(
|
||||
return cur;
|
||||
}
|
||||
|
||||
llm_graph_input_attn_k_dsa * llm_graph_context::build_attn_inp_k_dsa() const {
|
||||
const auto * mctx_cur = static_cast<const llama_kv_cache_dsa_context *>(mctx);
|
||||
|
||||
auto inp = std::make_unique<llm_graph_input_attn_k_dsa>(hparams, cparams, mctx_cur);
|
||||
|
||||
{
|
||||
inp->self_k_idxs_mla = mctx_cur->get_mla()->build_input_k_idxs(ctx0, ubatch);
|
||||
|
||||
inp->self_kq_mask_mla = build_attn_inp_kq_mask(ctx0, mctx_cur->get_mla(), ubatch, cparams);
|
||||
inp->self_kq_mask_mla_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_mla, GGML_TYPE_F16) : inp->self_kq_mask_mla;
|
||||
}
|
||||
|
||||
{
|
||||
inp->self_k_idxs_lid = mctx_cur->get_lid()->build_input_k_idxs(ctx0, ubatch);
|
||||
|
||||
inp->self_kq_mask_lid = build_attn_inp_kq_mask(ctx0, mctx_cur->get_lid(), ubatch, cparams);
|
||||
inp->self_kq_mask_lid_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_lid, GGML_TYPE_F16) : inp->self_kq_mask_lid;
|
||||
|
||||
inp->self_k_rot_lid = mctx_cur->get_lid()->build_input_k_rot(ctx0);
|
||||
}
|
||||
|
||||
return (llm_graph_input_attn_k_dsa *) res->add_input(std::move(inp));
|
||||
}
|
||||
|
||||
// TODO: maybe separate the inner implementation into a separate function
|
||||
// like with the non-sliding window equivalent
|
||||
// once sliding-window hybrid caches are a thing.
|
||||
@@ -2510,7 +2684,7 @@ llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const
|
||||
inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch);
|
||||
|
||||
inp->self_kq_mask = build_attn_inp_kq_mask(ctx0, mctx_cur->get_base(), ubatch, cparams);
|
||||
inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
|
||||
inp->self_kq_mask_cnv = inp->self_kq_mask;
|
||||
}
|
||||
|
||||
{
|
||||
@@ -2520,7 +2694,7 @@ llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const
|
||||
inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch);
|
||||
|
||||
inp->self_kq_mask_swa = build_attn_inp_kq_mask(ctx0, mctx_cur->get_swa(), ubatch, cparams);
|
||||
inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
|
||||
inp->self_kq_mask_swa_cnv = inp->self_kq_mask_swa;
|
||||
}
|
||||
|
||||
inp->self_k_rot = mctx_cur->get_base()->build_input_k_rot(ctx0);
|
||||
@@ -2689,7 +2863,7 @@ llm_graph_input_mem_hybrid_iswa * llm_graph_context::build_inp_mem_hybrid_iswa()
|
||||
inp_attn->self_v_idxs = attn_ctx->get_base()->build_input_v_idxs(ctx0, ubatch);
|
||||
|
||||
inp_attn->self_kq_mask = build_attn_inp_kq_mask(ctx0, attn_ctx->get_base(), ubatch, cparams);
|
||||
inp_attn->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp_attn->self_kq_mask, GGML_TYPE_F16) : inp_attn->self_kq_mask;
|
||||
inp_attn->self_kq_mask_cnv = inp_attn->self_kq_mask;
|
||||
}
|
||||
|
||||
{
|
||||
@@ -2697,7 +2871,7 @@ llm_graph_input_mem_hybrid_iswa * llm_graph_context::build_inp_mem_hybrid_iswa()
|
||||
inp_attn->self_v_idxs_swa = attn_ctx->get_swa()->build_input_v_idxs(ctx0, ubatch);
|
||||
|
||||
inp_attn->self_kq_mask_swa = build_attn_inp_kq_mask(ctx0, attn_ctx->get_swa(), ubatch, cparams);
|
||||
inp_attn->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp_attn->self_kq_mask_swa, GGML_TYPE_F16) : inp_attn->self_kq_mask_swa;
|
||||
inp_attn->self_kq_mask_swa_cnv = inp_attn->self_kq_mask_swa;
|
||||
}
|
||||
|
||||
auto inp = std::make_unique<llm_graph_input_mem_hybrid_iswa>(cparams, std::move(inp_attn), std::move(inp_rs), mctx_cur);
|
||||
|
||||
+87
-14
@@ -22,6 +22,7 @@ struct llama_layer;
|
||||
struct llama_memory_context_i;
|
||||
|
||||
class llama_kv_cache_context;
|
||||
class llama_kv_cache_dsa_context;
|
||||
class llama_kv_cache_iswa_context;
|
||||
class llama_memory_recurrent_context;
|
||||
class llama_memory_hybrid_context;
|
||||
@@ -121,6 +122,23 @@ public:
|
||||
const int64_t n_embd = 0;
|
||||
};
|
||||
|
||||
// similar to llm_graph_input_embd but with an additional hidden state input
|
||||
class llm_graph_input_embd_h : public llm_graph_input_i {
|
||||
public:
|
||||
llm_graph_input_embd_h(int64_t n_embd) : n_embd(n_embd) {}
|
||||
virtual ~llm_graph_input_embd_h() = default;
|
||||
|
||||
void set_input(const llama_ubatch * ubatch) override;
|
||||
|
||||
bool can_reuse(const llm_graph_params & params) override;
|
||||
|
||||
ggml_tensor * tokens = nullptr; // I32 [n_batch]
|
||||
ggml_tensor * embd = nullptr; // F32 [n_embd, n_batch]
|
||||
ggml_tensor * h = nullptr; // F32 [n_embd, n_batch]
|
||||
|
||||
const int64_t n_embd = 0;
|
||||
};
|
||||
|
||||
class llm_graph_input_pos : public llm_graph_input_i {
|
||||
public:
|
||||
llm_graph_input_pos(uint32_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {}
|
||||
@@ -274,10 +292,10 @@ public:
|
||||
ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
|
||||
|
||||
// n_tokens == n_batch
|
||||
ggml_tensor * self_kq_mask = nullptr; // F32 [n_tokens, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_tokens, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_swa = nullptr; // F32 [n_tokens, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_tokens, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask = nullptr; // F32/F16 [n_tokens, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_tokens, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_swa = nullptr; // F32/F16 [n_tokens, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_tokens, n_batch/n_stream, 1, n_stream]
|
||||
|
||||
const llama_hparams hparams;
|
||||
const llama_cparams cparams;
|
||||
@@ -307,8 +325,8 @@ public:
|
||||
ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
|
||||
ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
|
||||
|
||||
ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
|
||||
// note: assumes v_rot^2 == I
|
||||
ggml_tensor * self_k_rot = nullptr;
|
||||
@@ -347,8 +365,8 @@ public:
|
||||
|
||||
ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
|
||||
|
||||
ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
|
||||
const llama_hparams hparams;
|
||||
const llama_cparams cparams;
|
||||
@@ -356,6 +374,44 @@ public:
|
||||
const llama_kv_cache_context * mctx;
|
||||
};
|
||||
|
||||
class llm_graph_input_attn_k_dsa : public llm_graph_input_i {
|
||||
public:
|
||||
llm_graph_input_attn_k_dsa(
|
||||
const llama_hparams & hparams,
|
||||
const llama_cparams & cparams,
|
||||
const llama_kv_cache_dsa_context * mctx) :
|
||||
hparams(hparams),
|
||||
cparams(cparams),
|
||||
mctx(mctx) {
|
||||
}
|
||||
~llm_graph_input_attn_k_dsa() = default;
|
||||
|
||||
void set_input(const llama_ubatch * ubatch) override;
|
||||
|
||||
bool can_reuse(const llm_graph_params & params) override;
|
||||
|
||||
ggml_tensor * get_k_idxs_mla() const { return self_k_idxs_mla; }
|
||||
ggml_tensor * get_k_idxs_lid() const { return self_k_idxs_lid; }
|
||||
|
||||
ggml_tensor * get_kq_mask_mla() const { return self_kq_mask_mla_cnv; }
|
||||
ggml_tensor * get_kq_mask_lid() const { return self_kq_mask_lid; }
|
||||
|
||||
ggml_tensor * self_k_idxs_mla = nullptr; // I64 [n_batch]
|
||||
ggml_tensor * self_k_idxs_lid = nullptr; // I64 [n_batch]
|
||||
|
||||
ggml_tensor * self_kq_mask_mla = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_mla_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_lid = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_lid_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
|
||||
ggml_tensor * self_k_rot_lid = nullptr;
|
||||
|
||||
const llama_hparams hparams;
|
||||
const llama_cparams cparams;
|
||||
|
||||
const llama_kv_cache_dsa_context * mctx;
|
||||
};
|
||||
|
||||
class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
|
||||
public:
|
||||
llm_graph_input_attn_kv_iswa(
|
||||
@@ -385,10 +441,10 @@ public:
|
||||
ggml_tensor * self_k_idxs_swa = nullptr; // I64 [n_batch]
|
||||
ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
|
||||
|
||||
ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_swa = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_swa = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
|
||||
|
||||
ggml_tensor * self_k_rot = nullptr;
|
||||
ggml_tensor * self_v_rot = nullptr;
|
||||
@@ -411,8 +467,8 @@ public:
|
||||
|
||||
ggml_tensor * get_kq_mask_cross() const { return cross_kq_mask_cnv; }
|
||||
|
||||
ggml_tensor * cross_kq_mask = nullptr; // F32 [n_outputs_enc, n_batch, 1, 1]
|
||||
ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch, 1, 1]
|
||||
ggml_tensor * cross_kq_mask = nullptr; // F32/F16 [n_outputs_enc, n_batch, 1, 1]
|
||||
ggml_tensor * cross_kq_mask_cnv = nullptr; // F32/F16 [n_outputs_enc, n_batch, 1, 1]
|
||||
|
||||
const llama_cross * cross = nullptr;
|
||||
};
|
||||
@@ -956,6 +1012,23 @@ struct llm_graph_context {
|
||||
float kq_scale,
|
||||
int il) const;
|
||||
|
||||
llm_graph_input_attn_k_dsa * build_attn_inp_k_dsa() const;
|
||||
|
||||
ggml_tensor * build_attn(
|
||||
llm_graph_input_attn_k_dsa * inp,
|
||||
ggml_tensor * wo,
|
||||
ggml_tensor * wo_b,
|
||||
ggml_tensor * wo_s,
|
||||
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
|
||||
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
|
||||
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
|
||||
ggml_tensor * kq_b,
|
||||
ggml_tensor * sinks, // [n_head_q]
|
||||
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
|
||||
ggml_tensor * top_k, // [n_indexer_top_k, n_tokens]
|
||||
float kq_scale,
|
||||
int il) const;
|
||||
|
||||
llm_graph_input_attn_kv_iswa * build_attn_inp_kv_iswa() const;
|
||||
|
||||
// note: if k_cur or v_cur are not provided, they will not be stored in the memory
|
||||
|
||||
@@ -3,6 +3,7 @@
|
||||
#include "ggml.h" // for ggml_log_level
|
||||
|
||||
#include <string>
|
||||
#include <type_traits>
|
||||
#include <vector>
|
||||
|
||||
#ifdef __GNUC__
|
||||
@@ -40,6 +41,19 @@ struct no_init {
|
||||
no_init() = default;
|
||||
};
|
||||
|
||||
template <typename dst_t, typename src_t>
|
||||
static inline dst_t llama_cast(src_t v) {
|
||||
if constexpr (std::is_same_v<src_t, dst_t>) {
|
||||
return v;
|
||||
} else if constexpr (std::is_same_v<src_t, ggml_fp16_t> && std::is_same_v<dst_t, float>) {
|
||||
return ggml_fp16_to_fp32(v);
|
||||
} else if constexpr (std::is_same_v<src_t, float> && std::is_same_v<dst_t, ggml_fp16_t>) {
|
||||
return ggml_fp32_to_fp16(v);
|
||||
} else {
|
||||
static_assert(std::is_same_v<dst_t, void>, "unsupported type combination");
|
||||
}
|
||||
}
|
||||
|
||||
struct time_meas {
|
||||
time_meas(int64_t & t_acc, bool disable = false);
|
||||
~time_meas();
|
||||
|
||||
@@ -0,0 +1,261 @@
|
||||
#include "llama-kv-cache-dsa.h"
|
||||
|
||||
#include "llama-impl.h"
|
||||
#include "llama-batch.h"
|
||||
#include "llama-model.h"
|
||||
|
||||
#include <algorithm>
|
||||
#include <cassert>
|
||||
|
||||
//
|
||||
// llama_kv_cache_dsa
|
||||
//
|
||||
|
||||
llama_kv_cache_dsa::llama_kv_cache_dsa(
|
||||
const llama_model & model,
|
||||
ggml_type type_k,
|
||||
ggml_type type_v,
|
||||
bool v_trans,
|
||||
bool offload,
|
||||
bool unified,
|
||||
uint32_t kv_size,
|
||||
uint32_t n_seq_max,
|
||||
uint32_t n_pad,
|
||||
uint32_t n_swa,
|
||||
llama_swa_type swa_type,
|
||||
const layer_filter_cb & filter,
|
||||
const layer_reuse_cb & reuse) :
|
||||
hparams_lid(model.hparams), n_stream(unified ? 1 : n_seq_max) {
|
||||
|
||||
LLAMA_LOG_INFO("%s: creating main KV cache, size = %u cells\n", __func__, kv_size);
|
||||
|
||||
kv_mla = std::make_unique<llama_kv_cache>(
|
||||
model, model.hparams, type_k, type_v,
|
||||
v_trans, offload, unified, kv_size, n_seq_max, n_pad,
|
||||
n_swa, swa_type, filter, reuse);
|
||||
|
||||
// we use llama_kv_cache for caching indexer keys
|
||||
// by hand-tweaking some hparams we fool it to create
|
||||
// indexer key cache tensors with correct dimensions
|
||||
// https://github.com/ggml-org/llama.cpp/pull/21149#discussion_r3015940823
|
||||
|
||||
// DSA lightning indexer uses MQA with single key head
|
||||
std::fill(hparams_lid.n_head_kv_arr.begin(), hparams_lid.n_head_kv_arr.end(), 1);
|
||||
hparams_lid.n_embd_head_k_full = model.hparams.indexer_head_size;
|
||||
hparams_lid.rope_type = LLAMA_ROPE_TYPE_NEOX;
|
||||
|
||||
LLAMA_LOG_INFO("%s: creating indexer KV cache, size = %u cells\n", __func__, kv_size);
|
||||
|
||||
kv_lid = std::make_unique<llama_kv_cache>(
|
||||
model, hparams_lid, type_k, type_v,
|
||||
v_trans, offload, unified, kv_size, n_seq_max, n_pad,
|
||||
n_swa, swa_type, filter, reuse);
|
||||
}
|
||||
|
||||
void llama_kv_cache_dsa::clear(bool data) {
|
||||
kv_mla->clear(data);
|
||||
kv_lid->clear(data);
|
||||
}
|
||||
|
||||
bool llama_kv_cache_dsa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
|
||||
bool res = true;
|
||||
|
||||
res = res & kv_mla->seq_rm(seq_id, p0, p1);
|
||||
res = res & kv_lid->seq_rm(seq_id, p0, p1);
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
void llama_kv_cache_dsa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
|
||||
kv_mla->seq_cp(seq_id_src, seq_id_dst, p0, p1);
|
||||
kv_lid->seq_cp(seq_id_src, seq_id_dst, p0, p1);
|
||||
}
|
||||
|
||||
void llama_kv_cache_dsa::seq_keep(llama_seq_id seq_id) {
|
||||
kv_mla->seq_keep(seq_id);
|
||||
kv_lid->seq_keep(seq_id);
|
||||
}
|
||||
|
||||
void llama_kv_cache_dsa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
|
||||
kv_mla->seq_add(seq_id, p0, p1, shift);
|
||||
kv_lid->seq_add(seq_id, p0, p1, shift);
|
||||
}
|
||||
|
||||
void llama_kv_cache_dsa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
|
||||
kv_mla->seq_div(seq_id, p0, p1, d);
|
||||
kv_lid->seq_div(seq_id, p0, p1, d);
|
||||
}
|
||||
|
||||
llama_pos llama_kv_cache_dsa::seq_pos_min(llama_seq_id seq_id) const {
|
||||
return kv_mla->seq_pos_min(seq_id);
|
||||
}
|
||||
|
||||
llama_pos llama_kv_cache_dsa::seq_pos_max(llama_seq_id seq_id) const {
|
||||
return kv_mla->seq_pos_max(seq_id);
|
||||
}
|
||||
|
||||
std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache_dsa::memory_breakdown() const {
|
||||
std::map<ggml_backend_buffer_type_t, size_t> mb = kv_mla->memory_breakdown();
|
||||
for (const auto & buft_size : kv_lid->memory_breakdown()) {
|
||||
mb[buft_size.first] += buft_size.second;
|
||||
}
|
||||
return mb;
|
||||
}
|
||||
|
||||
llama_memory_context_ptr llama_kv_cache_dsa::init_batch(
|
||||
llama_batch_allocr & balloc,
|
||||
uint32_t n_ubatch,
|
||||
bool embd_all) {
|
||||
GGML_UNUSED(embd_all);
|
||||
|
||||
do {
|
||||
balloc.split_reset();
|
||||
|
||||
std::vector<llama_ubatch> ubatches;
|
||||
while (true) {
|
||||
auto ubatch = n_stream == 1 ? balloc.split_simple(n_ubatch) : balloc.split_equal(n_ubatch, true);
|
||||
|
||||
if (ubatch.n_tokens == 0) {
|
||||
break;
|
||||
}
|
||||
|
||||
ubatches.push_back(std::move(ubatch)); // NOLINT
|
||||
}
|
||||
|
||||
if (balloc.get_n_used() < balloc.get_n_tokens()) {
|
||||
// failed to find a suitable split
|
||||
break;
|
||||
}
|
||||
|
||||
auto sinfos_mla = kv_mla->prepare(ubatches);
|
||||
if (sinfos_mla.empty()) {
|
||||
break;
|
||||
}
|
||||
|
||||
auto sinfos_lid = kv_lid->prepare(ubatches);
|
||||
if (sinfos_lid.empty()) {
|
||||
break;
|
||||
}
|
||||
|
||||
assert(sinfos_mla.size() == sinfos_lid.size());
|
||||
|
||||
return std::make_unique<llama_kv_cache_dsa_context>(
|
||||
this, std::move(sinfos_mla), std::move(sinfos_lid), std::move(ubatches));
|
||||
} while (false);
|
||||
|
||||
return std::make_unique<llama_kv_cache_dsa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
|
||||
}
|
||||
|
||||
llama_memory_context_ptr llama_kv_cache_dsa::init_full() {
|
||||
return std::make_unique<llama_kv_cache_dsa_context>(this);
|
||||
}
|
||||
|
||||
llama_memory_context_ptr llama_kv_cache_dsa::init_update(llama_context * lctx, bool optimize) {
|
||||
return std::make_unique<llama_kv_cache_dsa_context>(this, lctx, optimize);
|
||||
}
|
||||
|
||||
bool llama_kv_cache_dsa::get_can_shift() const {
|
||||
return kv_mla->get_can_shift() &&
|
||||
kv_lid->get_can_shift() &&
|
||||
kv_mla->get_size() == kv_lid->get_size();
|
||||
}
|
||||
|
||||
void llama_kv_cache_dsa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
|
||||
kv_mla->state_write(io, seq_id, flags);
|
||||
kv_lid->state_write(io, seq_id, flags);
|
||||
}
|
||||
|
||||
void llama_kv_cache_dsa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
|
||||
kv_mla->state_read(io, seq_id, flags);
|
||||
kv_lid->state_read(io, seq_id, flags);
|
||||
}
|
||||
|
||||
llama_kv_cache * llama_kv_cache_dsa::get_mla() const {
|
||||
return kv_mla.get();
|
||||
}
|
||||
|
||||
llama_kv_cache * llama_kv_cache_dsa::get_lid() const {
|
||||
return kv_lid.get();
|
||||
}
|
||||
|
||||
//
|
||||
// llama_kv_cache_dsa_context
|
||||
//
|
||||
|
||||
llama_kv_cache_dsa_context::llama_kv_cache_dsa_context(llama_memory_status status) : status(status) {}
|
||||
|
||||
llama_kv_cache_dsa_context::llama_kv_cache_dsa_context(
|
||||
llama_kv_cache_dsa * kv) :
|
||||
ctx_mla(kv->get_mla()->init_full()),
|
||||
ctx_lid(kv->get_lid()->init_full()),
|
||||
status(llama_memory_status_combine(ctx_mla->get_status(), ctx_lid->get_status())) {
|
||||
}
|
||||
|
||||
llama_kv_cache_dsa_context::llama_kv_cache_dsa_context(
|
||||
llama_kv_cache_dsa * kv,
|
||||
llama_context * lctx,
|
||||
bool optimize) :
|
||||
ctx_mla(kv->get_mla()->init_update(lctx, optimize)),
|
||||
ctx_lid(kv->get_lid()->init_update(lctx, optimize)),
|
||||
status(llama_memory_status_combine(ctx_mla->get_status(), ctx_lid->get_status())) {
|
||||
}
|
||||
|
||||
llama_kv_cache_dsa_context::llama_kv_cache_dsa_context(
|
||||
llama_kv_cache_dsa * kv,
|
||||
slot_info_vec_t sinfos_mla,
|
||||
slot_info_vec_t sinfos_lid,
|
||||
std::vector<llama_ubatch> ubatches) :
|
||||
ubatches(std::move(ubatches)),
|
||||
// note: here we copy the ubatches. not sure if this is ideal
|
||||
ctx_mla(new llama_kv_cache_context(kv->get_mla(), std::move(sinfos_mla), this->ubatches)),
|
||||
ctx_lid(new llama_kv_cache_context(kv->get_lid(), std::move(sinfos_lid), this->ubatches)),
|
||||
status(llama_memory_status_combine(ctx_mla->get_status(), ctx_lid->get_status())) {
|
||||
}
|
||||
|
||||
llama_kv_cache_dsa_context:: ~llama_kv_cache_dsa_context() = default;
|
||||
|
||||
bool llama_kv_cache_dsa_context::next() {
|
||||
assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
|
||||
|
||||
ctx_mla->next();
|
||||
ctx_lid->next();
|
||||
|
||||
if (++i_next >= ubatches.size()) {
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
bool llama_kv_cache_dsa_context::apply() {
|
||||
assert(!llama_memory_status_is_fail(status));
|
||||
|
||||
bool res = true;
|
||||
|
||||
res = res & ctx_mla->apply();
|
||||
res = res & ctx_lid->apply();
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
llama_memory_status llama_kv_cache_dsa_context::get_status() const {
|
||||
return status;
|
||||
}
|
||||
|
||||
const llama_ubatch & llama_kv_cache_dsa_context::get_ubatch() const {
|
||||
assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
|
||||
|
||||
return ubatches[i_next];
|
||||
}
|
||||
|
||||
const llama_kv_cache_context * llama_kv_cache_dsa_context::get_mla() const {
|
||||
assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
|
||||
|
||||
return static_cast<const llama_kv_cache_context *>(ctx_mla.get());
|
||||
}
|
||||
|
||||
const llama_kv_cache_context * llama_kv_cache_dsa_context::get_lid() const {
|
||||
assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
|
||||
|
||||
return static_cast<const llama_kv_cache_context *>(ctx_lid.get());
|
||||
}
|
||||
@@ -0,0 +1,138 @@
|
||||
#pragma once
|
||||
|
||||
#include "llama-kv-cache.h"
|
||||
|
||||
#include <vector>
|
||||
|
||||
//
|
||||
// llama_kv_cache_dsa
|
||||
//
|
||||
|
||||
// utilizes two instances of llama_kv_cache:
|
||||
// - the first instance is for caching key tensors of the model,
|
||||
// - the second instance is for caching lightning indexer key tensors
|
||||
|
||||
class llama_kv_cache_dsa : public llama_memory_i {
|
||||
public:
|
||||
llama_kv_cache_dsa(
|
||||
const llama_model & model,
|
||||
ggml_type type_k,
|
||||
ggml_type type_v,
|
||||
bool v_trans,
|
||||
bool offload,
|
||||
bool unified,
|
||||
uint32_t kv_size,
|
||||
uint32_t n_seq_max,
|
||||
uint32_t n_pad,
|
||||
uint32_t n_swa,
|
||||
llama_swa_type swa_type,
|
||||
const layer_filter_cb & filter,
|
||||
const layer_reuse_cb & reuse);
|
||||
|
||||
~llama_kv_cache_dsa() = default;
|
||||
|
||||
//
|
||||
// llama_memory_i
|
||||
//
|
||||
|
||||
llama_memory_context_ptr init_batch(
|
||||
llama_batch_allocr & balloc,
|
||||
uint32_t n_ubatch,
|
||||
bool embd_all) override;
|
||||
|
||||
llama_memory_context_ptr init_full() override;
|
||||
|
||||
llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
|
||||
|
||||
bool get_can_shift() const override;
|
||||
|
||||
void clear(bool data) override;
|
||||
|
||||
bool seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) override;
|
||||
void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
|
||||
void seq_keep(llama_seq_id seq_id) override;
|
||||
void seq_add (llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) override;
|
||||
void seq_div (llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) override;
|
||||
|
||||
llama_pos seq_pos_min(llama_seq_id seq_id) const override;
|
||||
llama_pos seq_pos_max(llama_seq_id seq_id) const override;
|
||||
|
||||
std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const override;
|
||||
|
||||
// state write/load
|
||||
|
||||
void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
|
||||
void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
|
||||
|
||||
//
|
||||
// llama_kv_cache_dsa specific API
|
||||
//
|
||||
|
||||
llama_kv_cache * get_mla() const;
|
||||
llama_kv_cache * get_lid() const;
|
||||
|
||||
private:
|
||||
// we keep indexer KV cache hparams instance here as llama_kv_cache stores only reference to it
|
||||
llama_hparams hparams_lid;
|
||||
const uint32_t n_stream = 1;
|
||||
|
||||
std::unique_ptr<llama_kv_cache> kv_mla;
|
||||
std::unique_ptr<llama_kv_cache> kv_lid;
|
||||
};
|
||||
|
||||
class llama_kv_cache_dsa_context : public llama_memory_context_i {
|
||||
public:
|
||||
using slot_info_vec_t = llama_kv_cache::slot_info_vec_t;
|
||||
|
||||
// used for errors
|
||||
llama_kv_cache_dsa_context(llama_memory_status status);
|
||||
|
||||
// used to create a full-cache context
|
||||
llama_kv_cache_dsa_context(
|
||||
llama_kv_cache_dsa * kv);
|
||||
|
||||
// used to create an update context
|
||||
llama_kv_cache_dsa_context(
|
||||
llama_kv_cache_dsa * kv,
|
||||
llama_context * lctx,
|
||||
bool optimize);
|
||||
|
||||
// used to create a batch processing context from a batch
|
||||
llama_kv_cache_dsa_context(
|
||||
llama_kv_cache_dsa * kv,
|
||||
slot_info_vec_t sinfos_base,
|
||||
slot_info_vec_t sinfos_ik,
|
||||
std::vector<llama_ubatch> ubatches);
|
||||
|
||||
virtual ~llama_kv_cache_dsa_context();
|
||||
|
||||
//
|
||||
// llama_memory_context_i
|
||||
//
|
||||
|
||||
bool next() override;
|
||||
bool apply() override;
|
||||
|
||||
llama_memory_status get_status() const override;
|
||||
const llama_ubatch & get_ubatch() const override;
|
||||
|
||||
//
|
||||
// llama_kv_cache_dsa_context specific API
|
||||
//
|
||||
|
||||
const llama_kv_cache_context * get_mla() const;
|
||||
const llama_kv_cache_context * get_lid() const;
|
||||
|
||||
private:
|
||||
//llama_kv_cache_dsa * kv;
|
||||
|
||||
// the index of the next ubatch to process
|
||||
size_t i_next = 0;
|
||||
|
||||
std::vector<llama_ubatch> ubatches;
|
||||
|
||||
const llama_memory_context_ptr ctx_mla;
|
||||
const llama_memory_context_ptr ctx_lid;
|
||||
|
||||
const llama_memory_status status;
|
||||
};
|
||||
@@ -73,14 +73,14 @@ llama_kv_cache_iswa::llama_kv_cache_iswa(
|
||||
LLAMA_LOG_INFO("%s: creating non-SWA KV cache, size = %u cells\n", __func__, size_base);
|
||||
|
||||
kv_base = std::make_unique<llama_kv_cache>(
|
||||
model, type_k, type_v,
|
||||
model, hparams, type_k, type_v,
|
||||
v_trans, offload, unified, size_base, n_seq_max, n_pad,
|
||||
0, LLAMA_SWA_TYPE_NONE, filter_base, reuse);
|
||||
|
||||
LLAMA_LOG_INFO("%s: creating SWA KV cache, size = %u cells\n", __func__, size_swa);
|
||||
|
||||
kv_swa = std::make_unique<llama_kv_cache>(
|
||||
model, type_k, type_v,
|
||||
model, hparams, type_k, type_v,
|
||||
v_trans, offload, unified, size_swa, n_seq_max, n_pad,
|
||||
hparams.n_swa, hparams.swa_type, filter_swa, reuse);
|
||||
}
|
||||
|
||||
+40
-23
@@ -83,6 +83,7 @@ static ggml_tensor * ggml_mul_mat_aux(
|
||||
|
||||
llama_kv_cache::llama_kv_cache(
|
||||
const llama_model & model,
|
||||
const llama_hparams & hparams,
|
||||
ggml_type type_k,
|
||||
ggml_type type_v,
|
||||
bool v_trans,
|
||||
@@ -95,7 +96,7 @@ llama_kv_cache::llama_kv_cache(
|
||||
llama_swa_type swa_type,
|
||||
const layer_filter_cb & filter,
|
||||
const layer_reuse_cb & reuse) :
|
||||
model(model), hparams(model.hparams), v_trans(v_trans),
|
||||
model(model), hparams(hparams), v_trans(v_trans),
|
||||
n_seq_max(n_seq_max), n_stream(unified ? 1 : n_seq_max), n_pad(n_pad), n_swa(n_swa), swa_type(swa_type) {
|
||||
|
||||
GGML_ASSERT(kv_size % n_pad == 0);
|
||||
@@ -257,7 +258,7 @@ llama_kv_cache::llama_kv_cache(
|
||||
// allocate tensors and initialize the buffers to avoid NaNs in the padding
|
||||
for (auto & [buft, ctx] : ctx_map) {
|
||||
ggml_backend_buffer_t buf;
|
||||
if (model.hparams.no_alloc) {
|
||||
if (hparams.no_alloc) {
|
||||
buf = ggml_backend_buft_alloc_buffer(buft, /*size =*/ 0); // dummy buffer
|
||||
for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != nullptr; t = ggml_get_next_tensor(ctx.get(), t)) {
|
||||
t->buffer = buf; // set dummy buffer for KV cache so that the backend scheduler won't try to allocate it
|
||||
@@ -297,6 +298,11 @@ llama_kv_cache::llama_kv_cache(
|
||||
ggml_is_quantized(type_k) &&
|
||||
hparams.n_embd_head_k() % 64 == 0;
|
||||
|
||||
// always create Hadamard rotation tensors for DeepSeek V3.2 DSA lightning indexer
|
||||
if (model.arch == LLM_ARCH_DEEPSEEK32 && hparams.n_embd_head_k_full == hparams.indexer_head_size) {
|
||||
attn_rot_k = true;
|
||||
}
|
||||
|
||||
attn_rot_v =
|
||||
!attn_rot_disable &&
|
||||
n_embd_head_v_all > 0 &&
|
||||
@@ -1439,8 +1445,8 @@ struct args_set_input_kq_mask {
|
||||
int64_t n_tps;
|
||||
};
|
||||
|
||||
template<bool causal, bool swa, bool is_2d, bool alibi>
|
||||
static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float * data) {
|
||||
template<typename T, bool causal, bool swa, bool is_2d, bool alibi>
|
||||
static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, T * data) {
|
||||
//const auto & hparams = args.hparams;
|
||||
const auto & ubatch = args.ubatch;
|
||||
|
||||
@@ -1454,6 +1460,9 @@ static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float *
|
||||
const int64_t n_stream = args.n_stream;
|
||||
const int64_t n_tps = args.n_tps;
|
||||
|
||||
const T mask_keep = llama_cast<T>(0.0f);
|
||||
const T mask_drop = llama_cast<T>(-INFINITY);
|
||||
|
||||
// the min position in the batch for each sequence
|
||||
llama_pos seq_pos_min[LLAMA_MAX_SEQ];
|
||||
std::fill(seq_pos_min, seq_pos_min + LLAMA_MAX_SEQ, INT32_MAX);
|
||||
@@ -1572,46 +1581,55 @@ static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float *
|
||||
}
|
||||
|
||||
if (alibi) {
|
||||
data[idst + j] = -std::abs(p0 - p1);
|
||||
data[idst + j] = llama_cast<T>(static_cast<float>(-std::abs(p0 - p1)));
|
||||
} else {
|
||||
data[idst + j] = 0.0f;
|
||||
data[idst + j] = mask_keep;
|
||||
}
|
||||
|
||||
continue;
|
||||
skip:
|
||||
data[idst + j] = -INFINITY;
|
||||
data[idst + j] = mask_drop;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template<bool causal, bool swa, bool is_2d>
|
||||
static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float * data) {
|
||||
template<typename T, bool causal, bool swa, bool is_2d>
|
||||
static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, T * data) {
|
||||
const bool alibi = args.hparams.use_alibi;
|
||||
if (alibi) {
|
||||
set_input_kq_mask_impl<causal, swa, is_2d, true> (args, data);
|
||||
set_input_kq_mask_impl<T, causal, swa, is_2d, true> (args, data);
|
||||
} else {
|
||||
set_input_kq_mask_impl<causal, swa, is_2d, false>(args, data);
|
||||
set_input_kq_mask_impl<T, causal, swa, is_2d, false>(args, data);
|
||||
}
|
||||
}
|
||||
|
||||
template<bool causal, bool swa>
|
||||
static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float * data) {
|
||||
template<typename T, bool causal, bool swa>
|
||||
static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, T * data) {
|
||||
const bool is_2d = args.ubatch->is_pos_2d();
|
||||
if (is_2d) {
|
||||
set_input_kq_mask_impl<causal, swa, true> (args, data);
|
||||
set_input_kq_mask_impl<T, causal, swa, true> (args, data);
|
||||
} else {
|
||||
set_input_kq_mask_impl<causal, swa, false>(args, data);
|
||||
set_input_kq_mask_impl<T, causal, swa, false>(args, data);
|
||||
}
|
||||
}
|
||||
|
||||
template<bool causal>
|
||||
static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float * data) {
|
||||
template<typename T, bool causal>
|
||||
static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, T * data) {
|
||||
const bool swa = args.swa_type != LLAMA_SWA_TYPE_NONE;
|
||||
if (swa) {
|
||||
set_input_kq_mask_impl<causal, true> (args, data);
|
||||
set_input_kq_mask_impl<T, causal, true> (args, data);
|
||||
} else {
|
||||
set_input_kq_mask_impl<causal, false>(args, data);
|
||||
set_input_kq_mask_impl<T, causal, false>(args, data);
|
||||
}
|
||||
}
|
||||
|
||||
template<typename T>
|
||||
static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, T * data, bool causal_attn) {
|
||||
if (causal_attn) {
|
||||
set_input_kq_mask_impl<T, true> (args, data);
|
||||
} else {
|
||||
set_input_kq_mask_impl<T, false>(args, data);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -1619,7 +1637,6 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u
|
||||
const uint32_t n_tokens = ubatch->n_tokens;
|
||||
|
||||
GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
|
||||
float * data = (float *) dst->data;
|
||||
|
||||
const int64_t n_kv = dst->ne[0];
|
||||
const int64_t n_stream = dst->ne[3]; // num streams in the current ubatch
|
||||
@@ -1643,10 +1660,10 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u
|
||||
/*.n_tps =*/ n_tps,
|
||||
};
|
||||
|
||||
if (causal_attn) {
|
||||
set_input_kq_mask_impl<true> (args, data);
|
||||
if (dst->type == GGML_TYPE_F16) {
|
||||
set_input_kq_mask_impl<ggml_fp16_t>(args, (ggml_fp16_t *) dst->data, causal_attn);
|
||||
} else {
|
||||
set_input_kq_mask_impl<false>(args, data);
|
||||
set_input_kq_mask_impl<float>(args, (float *) dst->data, causal_attn);
|
||||
}
|
||||
|
||||
//const int64_t t_end = ggml_time_us();
|
||||
|
||||
@@ -93,8 +93,12 @@ public:
|
||||
|
||||
using slot_info_vec_t = std::vector<slot_info>;
|
||||
|
||||
// TODO: refactor the memory instances to not depend on `llama_model`
|
||||
// instead pass all necessary info (e.g. hparams, dev layers, arch, etc.) directly
|
||||
// likely through `struct llama_memory_params`
|
||||
llama_kv_cache(
|
||||
const llama_model & model,
|
||||
const llama_hparams & hparams,
|
||||
ggml_type type_k,
|
||||
ggml_type type_v,
|
||||
bool v_trans,
|
||||
|
||||
@@ -33,6 +33,7 @@ llama_memory_hybrid::llama_memory_hybrid(
|
||||
hparams(model.hparams),
|
||||
mem_attn(new llama_kv_cache(
|
||||
model,
|
||||
model.hparams,
|
||||
type_k,
|
||||
type_v,
|
||||
v_trans,
|
||||
|
||||
+25
-1
@@ -10,6 +10,7 @@
|
||||
|
||||
#include "llama-kv-cache.h"
|
||||
#include "llama-kv-cache-iswa.h"
|
||||
#include "llama-kv-cache-dsa.h"
|
||||
#include "llama-memory-hybrid.h"
|
||||
#include "llama-memory-hybrid-iswa.h"
|
||||
#include "llama-memory-recurrent.h"
|
||||
@@ -57,6 +58,7 @@
|
||||
#include "models/deepseek.cpp"
|
||||
#include "models/deepseek2.cpp"
|
||||
#include "models/deepseek2ocr.cpp"
|
||||
#include "models/deepseek32.cpp"
|
||||
#include "models/delta-net-base.cpp"
|
||||
#include "models/dots1.cpp"
|
||||
#include "models/dream.cpp"
|
||||
@@ -303,6 +305,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
|
||||
return new llama_model_deepseek2(params);
|
||||
case LLM_ARCH_DEEPSEEK2OCR:
|
||||
return new llama_model_deepseek2ocr(params);
|
||||
case LLM_ARCH_DEEPSEEK32:
|
||||
return new llama_model_deepseek32(params);
|
||||
case LLM_ARCH_GLM_DSA:
|
||||
return new llama_model_glm_dsa(params);
|
||||
case LLM_ARCH_MISTRAL4:
|
||||
@@ -910,6 +914,7 @@ const char * llm_type_name(llm_type type) {
|
||||
case LLM_TYPE_310B_A15B: return "310B.A15B";
|
||||
case LLM_TYPE_355B_A32B: return "355B.A32B";
|
||||
case LLM_TYPE_397B_A17B: return "397B.A17B";
|
||||
case LLM_TYPE_685B_A37B: return "685B.A37B";
|
||||
case LLM_TYPE_744B_A40B: return "744B.A40B";
|
||||
case LLM_TYPE_E2B: return "E2B";
|
||||
case LLM_TYPE_E4B: return "E4B";
|
||||
@@ -1901,7 +1906,7 @@ void llama_model::print_info() const {
|
||||
LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale);
|
||||
}
|
||||
|
||||
if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_DEEPSEEK2OCR || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_MISTRAL4) {
|
||||
if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_DEEPSEEK2OCR || arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_MISTRAL4) {
|
||||
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
|
||||
LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q);
|
||||
LLAMA_LOG_INFO("%s: n_lora_kv = %d\n", __func__, hparams.n_lora_kv);
|
||||
@@ -2089,6 +2094,23 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
|
||||
{
|
||||
res = nullptr;
|
||||
} break;
|
||||
case LLM_ARCH_DEEPSEEK32:
|
||||
{
|
||||
res = new llama_kv_cache_dsa(
|
||||
*this,
|
||||
params.type_k,
|
||||
params.type_v,
|
||||
!cparams.flash_attn,
|
||||
cparams.offload_kqv,
|
||||
cparams.kv_unified,
|
||||
cparams.n_ctx_seq,
|
||||
cparams.n_seq_max,
|
||||
1,
|
||||
hparams.n_swa,
|
||||
hparams.swa_type,
|
||||
nullptr,
|
||||
nullptr);
|
||||
} break;
|
||||
// Models that need standard caching should rely on recurrent/hybrid
|
||||
// checks
|
||||
default:
|
||||
@@ -2215,6 +2237,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
|
||||
|
||||
res = new llama_kv_cache(
|
||||
*this,
|
||||
hparams,
|
||||
params.type_k,
|
||||
params.type_v,
|
||||
!cparams.flash_attn,
|
||||
@@ -2404,6 +2427,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
|
||||
case LLM_ARCH_DEEPSEEK:
|
||||
case LLM_ARCH_DEEPSEEK2:
|
||||
case LLM_ARCH_DEEPSEEK2OCR:
|
||||
case LLM_ARCH_DEEPSEEK32:
|
||||
case LLM_ARCH_PLM:
|
||||
case LLM_ARCH_CHATGLM:
|
||||
case LLM_ARCH_GRANITE:
|
||||
|
||||
@@ -137,6 +137,7 @@ enum llm_type {
|
||||
LLM_TYPE_310B_A15B, // /MiMo-V2-Flash
|
||||
LLM_TYPE_355B_A32B, // GLM-4.5
|
||||
LLM_TYPE_397B_A17B, // Qwen3.5
|
||||
LLM_TYPE_685B_A37B, // DeepSeek V3.2
|
||||
LLM_TYPE_744B_A40B, // GLM-5
|
||||
LLM_TYPE_E2B,
|
||||
LLM_TYPE_E4B,
|
||||
|
||||
@@ -11,6 +11,7 @@
|
||||
#include "llama-grammar.cpp"
|
||||
#include "llama-sampler.cpp"
|
||||
#include "llama-kv-cache.cpp"
|
||||
#include "llama-kv-cache-dsa.cpp"
|
||||
#include "llama-kv-cache-iswa.cpp"
|
||||
#include "llama-memory-hybrid.cpp"
|
||||
#include "llama-memory-hybrid-iswa.cpp"
|
||||
|
||||
@@ -0,0 +1,503 @@
|
||||
#include "models.h"
|
||||
|
||||
#include "llama-kv-cache.h"
|
||||
#include "llama-kv-cache-dsa.h"
|
||||
|
||||
void llama_model_deepseek32::load_arch_hparams(llama_model_loader & ml) {
|
||||
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
|
||||
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
|
||||
hparams.f_norm_eps = 1e-6; // eps for layer norm
|
||||
ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
|
||||
|
||||
// MoE parameters
|
||||
ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert);
|
||||
ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used);
|
||||
ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
|
||||
ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead, false);
|
||||
ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false);
|
||||
ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false);
|
||||
|
||||
// deepseek MLA parameters
|
||||
ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);
|
||||
ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv);
|
||||
ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA, hparams.n_embd_head_k_mla_impl, false);
|
||||
ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false);
|
||||
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
|
||||
ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
|
||||
|
||||
// DSA parameters
|
||||
ml.get_key(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, hparams.indexer_n_head);
|
||||
ml.get_key(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, hparams.indexer_head_size);
|
||||
ml.get_key(LLM_KV_ATTENTION_INDEXER_TOP_K, hparams.indexer_top_k);
|
||||
|
||||
// Expert gating function
|
||||
ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func);
|
||||
|
||||
if (ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, 0.0f)) {
|
||||
// [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
|
||||
// cancel the factor from the convert script
|
||||
hparams.rope_yarn_log_mul /= 0.1f;
|
||||
}
|
||||
|
||||
// NextN/MTP parameters
|
||||
ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false);
|
||||
GGML_ASSERT(hparams.nextn_predict_layers < hparams.n_layer && "nextn_predict_layers must be < n_layer");
|
||||
|
||||
// TODO: when MTP is implemented, this should probably be updated if needed
|
||||
hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
|
||||
|
||||
switch (hparams.n_layer) {
|
||||
case 62: type = LLM_TYPE_685B_A37B; break;
|
||||
default: type = LLM_TYPE_UNKNOWN;
|
||||
}
|
||||
}
|
||||
|
||||
void llama_model_deepseek32::load_arch_tensors(llama_model_loader &) {
|
||||
LLAMA_LOAD_LOCALS;
|
||||
const bool is_mla = hparams.is_mla();
|
||||
if (!is_mla) {
|
||||
throw std::runtime_error("DEEPSEEK32 architecture requires MLA");
|
||||
}
|
||||
|
||||
// note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
|
||||
const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
|
||||
const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
|
||||
|
||||
const int64_t n_embd_head_qk_rope = hparams.n_rot();
|
||||
const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
|
||||
|
||||
const int64_t q_lora_rank = hparams.n_lora_q;
|
||||
const int64_t kv_lora_rank = hparams.n_lora_kv;
|
||||
|
||||
const int64_t n_ff_exp = hparams.n_ff_exp;
|
||||
const int64_t n_expert_shared = hparams.n_expert_shared;
|
||||
|
||||
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
|
||||
|
||||
// output
|
||||
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
|
||||
// try to load output.weight, if not found, use token_embd (tied embeddings)
|
||||
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
|
||||
if (!output) {
|
||||
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
|
||||
}
|
||||
|
||||
for (int i = 0; i < n_layer; ++i) {
|
||||
int flags = 0;
|
||||
if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
|
||||
// skip all tensors in the NextN layers
|
||||
// TODO @ngxson : TENSOR_NOT_REQUIRED was a hack, need to remove it later
|
||||
flags |= TENSOR_SKIP | TENSOR_NOT_REQUIRED;
|
||||
}
|
||||
|
||||
auto & layer = layers[i];
|
||||
|
||||
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);
|
||||
layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, flags);
|
||||
layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, flags);
|
||||
|
||||
layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, flags);
|
||||
layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, flags);
|
||||
|
||||
layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, flags);
|
||||
|
||||
// note: only old legacy GGUF files will have the unsplit wkv_b tensor in
|
||||
layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, flags);
|
||||
layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, flags);
|
||||
|
||||
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, flags);
|
||||
|
||||
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
|
||||
|
||||
// DSA indexer
|
||||
layer.indexer_k_norm = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "weight", i), {hparams.indexer_head_size}, flags);
|
||||
layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "bias", i), {hparams.indexer_head_size}, flags);
|
||||
layer.indexer_proj = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ, "weight", i), {n_embd, hparams.indexer_n_head}, flags);
|
||||
layer.indexer_attn_k = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K, "weight", i), {n_embd, hparams.indexer_head_size}, flags);
|
||||
layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags);
|
||||
if (i < (int) hparams.n_layer_dense_lead) {
|
||||
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, flags);
|
||||
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, flags);
|
||||
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, flags);
|
||||
} else {
|
||||
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, flags);
|
||||
layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
|
||||
|
||||
if (n_expert == 0) {
|
||||
throw std::runtime_error("n_expert must be > 0");
|
||||
}
|
||||
if (n_expert_used == 0) {
|
||||
throw std::runtime_error("n_expert_used must be > 0");
|
||||
}
|
||||
|
||||
// MoE branch
|
||||
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, flags);
|
||||
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, flags);
|
||||
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, flags);
|
||||
|
||||
// Shared expert branch
|
||||
layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
|
||||
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_exp * n_expert_shared, n_embd}, flags);
|
||||
layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
|
||||
}
|
||||
|
||||
// NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
|
||||
if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
|
||||
layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
|
||||
layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
|
||||
layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
|
||||
|
||||
// Optional tensors
|
||||
layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
|
||||
layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
|
||||
layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
std::unique_ptr<llm_graph_context> llama_model_deepseek32::build_arch_graph(const llm_graph_params & params) const {
|
||||
return std::make_unique<graph>(*this, params);
|
||||
}
|
||||
|
||||
llama_model_deepseek32::graph::graph(const llama_model & model, const llm_graph_params & params) :
|
||||
llm_graph_context(params) {
|
||||
const bool is_mla = hparams.is_mla();
|
||||
GGML_ASSERT(is_mla);
|
||||
|
||||
// note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
|
||||
const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
|
||||
const int64_t n_embd_head_v = hparams.n_embd_head_v_mla();
|
||||
GGML_UNUSED(n_embd_head_v);
|
||||
|
||||
const int64_t n_embd_head_qk_rope = hparams.n_rot();
|
||||
const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
|
||||
|
||||
const int64_t n_indexer_head = hparams.indexer_n_head;
|
||||
const int64_t n_embd_indexer_head = hparams.indexer_head_size;
|
||||
const int64_t n_embd_indexer_head_rope = hparams.n_rot();
|
||||
const int64_t n_embd_indexer_head_nope = n_embd_indexer_head - n_embd_indexer_head_rope;
|
||||
const uint32_t n_indexer_top_k = hparams.indexer_top_k;
|
||||
|
||||
const uint32_t kv_lora_rank = hparams.n_lora_kv;
|
||||
|
||||
// We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
|
||||
// See https://github.com/ggml-org/llama.cpp/discussions/7416 for detailed explanation.
|
||||
// And also: https://github.com/ggml-org/llama.cpp/pull/17945 [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
|
||||
|
||||
// first cancel the adjustment from llama_hparams::yarn_attn_factor_adjust to get the original attn_factor
|
||||
GGML_ASSERT(ext_factor >= 0.0f);
|
||||
const float attn_factor_org = attn_factor * (1.0f + 0.1f * logf(1.0f / freq_scale));
|
||||
|
||||
// use the original attn_factor to pre-scale the kq_scale
|
||||
const float mscale = attn_factor_org * (1.0f + 0.1f * hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
|
||||
const float kq_scale = 1.0f * mscale * mscale / sqrtf(float(n_embd_head_k));
|
||||
|
||||
ggml_tensor * cur;
|
||||
ggml_tensor * inpL;
|
||||
|
||||
// {n_embd, n_tokens}
|
||||
inpL = build_inp_embd(model.tok_embd);
|
||||
|
||||
// inp_pos - contains the positions
|
||||
ggml_tensor * inp_pos = build_inp_pos();
|
||||
|
||||
llm_graph_input_attn_k_dsa * inp_attn_dsa = build_attn_inp_k_dsa();
|
||||
|
||||
ggml_tensor * inp_out_ids = build_inp_out_ids();
|
||||
|
||||
int effective_n_layers = hparams.n_layer - hparams.nextn_predict_layers;
|
||||
for (int il = 0; il < effective_n_layers; ++il) {
|
||||
ggml_tensor * inpSA = inpL;
|
||||
|
||||
// norm
|
||||
cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
|
||||
cb(cur, "attn_norm", il);
|
||||
|
||||
// self_attention
|
||||
{
|
||||
ggml_tensor * qr = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
|
||||
cb(qr, "qr", il);
|
||||
|
||||
qr = build_norm(qr, model.layers[il].attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
|
||||
cb(qr, "qr", il);
|
||||
|
||||
ggml_tensor * top_k = nullptr;
|
||||
|
||||
// lightning indexer
|
||||
{
|
||||
ggml_tensor * indexer_q = ggml_mul_mat(ctx0, model.layers[il].indexer_attn_q_b, qr);
|
||||
cb(indexer_q, "indexer_q", il);
|
||||
|
||||
// split into {n_embd_indexer_head_rope, n_indexer_head, n_tokens}
|
||||
ggml_tensor * indexer_q_pe =
|
||||
ggml_view_3d(ctx0, indexer_q, n_embd_indexer_head_rope, n_indexer_head, n_tokens,
|
||||
ggml_row_size(indexer_q->type, n_embd_indexer_head),
|
||||
ggml_row_size(indexer_q->type, n_embd_indexer_head) * n_indexer_head, 0);
|
||||
cb(indexer_q_pe, "indexer_q_pe", il);
|
||||
|
||||
// and {n_embd_indexer_head_nope, n_indexer_head, n_tokens}
|
||||
ggml_tensor * indexer_q_nope =
|
||||
ggml_view_3d(ctx0, indexer_q, n_embd_indexer_head_nope, n_indexer_head, n_tokens,
|
||||
ggml_row_size(indexer_q->type, n_embd_indexer_head),
|
||||
ggml_row_size(indexer_q->type, n_embd_indexer_head) * n_indexer_head,
|
||||
ggml_row_size(indexer_q->type, n_embd_indexer_head_nope));
|
||||
cb(indexer_q_nope, "indexer_q_nope", il);
|
||||
|
||||
indexer_q_pe = ggml_rope_ext(ctx0, indexer_q_pe, inp_pos, nullptr, n_rot,
|
||||
LLAMA_ROPE_TYPE_NEOX, n_ctx_orig, freq_base, freq_scale,
|
||||
ext_factor, attn_factor, beta_fast, beta_slow);
|
||||
cb(indexer_q_pe, "indexer_q_pe", il);
|
||||
|
||||
// {n_embd_indexer_head_rope + n_embd_indexer_head_nope, n_head, n_tokens}
|
||||
indexer_q = ggml_concat(ctx0, indexer_q_pe, indexer_q_nope, 0);
|
||||
cb(indexer_q, "indexer_q", il);
|
||||
|
||||
ggml_tensor * indexer_k = ggml_mul_mat(ctx0, model.layers[il].indexer_attn_k, cur);
|
||||
cb(indexer_k, "indexer_k", il);
|
||||
|
||||
indexer_k = build_norm(indexer_k, model.layers[il].indexer_k_norm, model.layers[il].indexer_k_norm_b, LLM_NORM, il);
|
||||
cb(indexer_k, "indexer_k", il);
|
||||
|
||||
// split into {n_embd_indexer_head_rope, 1, n_tokens}
|
||||
ggml_tensor * indexer_k_pe =
|
||||
ggml_view_3d(ctx0, indexer_k, n_embd_indexer_head_rope, 1, n_tokens,
|
||||
ggml_row_size(indexer_k->type, n_embd_indexer_head),
|
||||
ggml_row_size(indexer_k->type, n_embd_indexer_head) * 1, 0);
|
||||
cb(indexer_k_pe, "indexer_k_pe", il);
|
||||
|
||||
// and {n_embd_indexer_head_nope, 1, n_tokens}
|
||||
ggml_tensor * indexer_k_nope =
|
||||
ggml_view_3d(ctx0, indexer_k, n_embd_indexer_head_nope, 1, n_tokens,
|
||||
ggml_row_size(indexer_k->type, n_embd_indexer_head),
|
||||
ggml_row_size(indexer_k->type, n_embd_indexer_head) * 1,
|
||||
ggml_row_size(indexer_k->type, n_embd_indexer_head_nope));
|
||||
cb(indexer_k_nope, "indexer_k_nope", il);
|
||||
|
||||
indexer_k_pe = ggml_rope_ext(ctx0, indexer_k_pe, inp_pos, nullptr, n_rot,
|
||||
LLAMA_ROPE_TYPE_NEOX, n_ctx_orig, freq_base, freq_scale,
|
||||
ext_factor, attn_factor, beta_fast, beta_slow);
|
||||
cb(indexer_k_pe, "indexer_k_pe", il);
|
||||
|
||||
// {n_embd_indexer_head_rope + n_embd_indexer_head_nope, 1, n_tokens}
|
||||
indexer_k = ggml_concat(ctx0, indexer_k_pe, indexer_k_nope, 0);
|
||||
cb(indexer_k, "indexer_k", il);
|
||||
|
||||
// perform Hadamard transform on indexer q and k
|
||||
indexer_q = ggml_mul_mat(ctx0, inp_attn_dsa->self_k_rot_lid, indexer_q);
|
||||
cb(indexer_q, "indexer_q", il);
|
||||
indexer_k = ggml_mul_mat(ctx0, inp_attn_dsa->self_k_rot_lid, indexer_k);
|
||||
cb(indexer_k, "indexer_k", il);
|
||||
|
||||
// store indexer keys to KV cache
|
||||
const auto * mctx_lid = inp_attn_dsa->mctx->get_lid();
|
||||
const auto & k_idxs_lid = inp_attn_dsa->get_k_idxs_lid();
|
||||
ggml_build_forward_expand(gf, mctx_lid->cpy_k(ctx0, indexer_k, k_idxs_lid, il));
|
||||
|
||||
// prepare indexer weights
|
||||
ggml_tensor * indexer_weights = ggml_mul_mat(ctx0, model.layers[il].indexer_proj, cur);
|
||||
cb(indexer_weights, "indexer_weights", il);
|
||||
|
||||
// get cached indexer keys
|
||||
indexer_k = mctx_lid->get_k(ctx0, il);
|
||||
|
||||
// split the batch into streams if needed
|
||||
const auto n_stream = indexer_k->ne[3];
|
||||
indexer_q = ggml_view_4d(ctx0, indexer_q, indexer_q->ne[0], indexer_q->ne[1], indexer_q->ne[2]/n_stream, n_stream, indexer_q->nb[1], indexer_q->nb[2], indexer_q->nb[3]/n_stream, 0);
|
||||
indexer_weights = ggml_view_4d(ctx0, indexer_weights, indexer_weights->ne[0], indexer_weights->ne[1]/n_stream, indexer_weights->ne[2], n_stream, indexer_weights->nb[1], indexer_weights->nb[2]/n_stream, indexer_weights->nb[3]/n_stream, 0);
|
||||
|
||||
// calculate indexer kq
|
||||
indexer_q = ggml_permute(ctx0, indexer_q, 0, 2, 1, 3);
|
||||
cb(indexer_q, "indexer_q", il);
|
||||
indexer_k = ggml_permute(ctx0, indexer_k, 0, 2, 1, 3);
|
||||
cb(indexer_k, "indexer_k", il);
|
||||
|
||||
ggml_tensor * indexer_kq = ggml_mul_mat(ctx0, indexer_k, indexer_q);
|
||||
cb(indexer_kq, "indexer_kq", il);
|
||||
|
||||
// ReLU requires contiguous tensors
|
||||
indexer_kq = ggml_cont(ctx0, ggml_permute(ctx0, indexer_kq, 2, 1, 0, 3));
|
||||
cb(indexer_kq, "indexer_kq", il);
|
||||
|
||||
// apply ReLU
|
||||
ggml_tensor * indexer_score = ggml_relu(ctx0, indexer_kq);
|
||||
cb(indexer_score, "indexer_score", il);
|
||||
|
||||
// pre-scale weights to avoid scaling operations on huge indexer_score tensor
|
||||
indexer_weights = ggml_scale(ctx0, indexer_weights, 1.0f / sqrtf(float(n_embd_indexer_head * n_indexer_head)));
|
||||
cb(indexer_weights, "indexer_weights", il);
|
||||
|
||||
// multiply scores by indexer weights
|
||||
indexer_score = ggml_mul(ctx0, indexer_score, indexer_weights);
|
||||
cb(indexer_score, "indexer_score", il);
|
||||
|
||||
// sum by q n_indexer_head dimension
|
||||
indexer_score = ggml_sum_rows(ctx0, indexer_score);
|
||||
cb(indexer_score, "indexer_score", il);
|
||||
|
||||
// permute result to match KQ mask
|
||||
indexer_score = ggml_cont(ctx0, ggml_permute(ctx0, indexer_score, 2, 1, 0, 3));
|
||||
cb(indexer_score, "indexer_score", il);
|
||||
|
||||
// mask indexer scores
|
||||
ggml_tensor * indexer_kq_mask = inp_attn_dsa->get_kq_mask_lid();
|
||||
indexer_score = ggml_add(ctx0, indexer_score, indexer_kq_mask);
|
||||
cb(indexer_score, "indexer_score", il);
|
||||
|
||||
// get indices of top k indexer scores
|
||||
uint32_t n_top_k = indexer_score->ne[0] < n_indexer_top_k ? indexer_score->ne[0] : n_indexer_top_k;
|
||||
top_k = ggml_cont(ctx0, ggml_top_k(ctx0, indexer_score, n_top_k));
|
||||
cb(top_k, "top_k", il);
|
||||
}
|
||||
|
||||
ggml_tensor * q = ggml_mul_mat(ctx0, model.layers[il].wq_b, qr);
|
||||
cb(q, "q", il);
|
||||
|
||||
// split into {n_embd_head_qk_nope, n_head, n_tokens}
|
||||
ggml_tensor * q_nope =
|
||||
ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
|
||||
ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
|
||||
cb(q_nope, "q_nope", il);
|
||||
|
||||
// and {n_embd_head_qk_rope, n_head, n_tokens}
|
||||
ggml_tensor * q_pe = ggml_view_3d(
|
||||
ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
|
||||
ggml_row_size(q->type, n_embd_head_k) * n_head, ggml_row_size(q->type, n_embd_head_qk_nope));
|
||||
cb(q_pe, "q_pe", il);
|
||||
|
||||
ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
|
||||
cb(kv_cmpr_pe, "kv_cmpr_pe", il);
|
||||
|
||||
// split into {kv_lora_rank, n_tokens}
|
||||
ggml_tensor * kv_cmpr =
|
||||
ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
|
||||
ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
|
||||
cb(kv_cmpr, "kv_cmpr", il);
|
||||
|
||||
// and {n_embd_head_qk_rope, 1, n_tokens}
|
||||
ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
|
||||
ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
|
||||
ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
|
||||
ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
|
||||
cb(k_pe, "k_pe", il);
|
||||
|
||||
q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
||||
ext_factor, attn_factor, beta_fast, beta_slow);
|
||||
cb(q_pe, "q_pe", il);
|
||||
|
||||
k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
||||
ext_factor, attn_factor, beta_fast, beta_slow);
|
||||
cb(k_pe, "k_pe", il);
|
||||
|
||||
kv_cmpr = build_norm(kv_cmpr, model.layers[il].attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
|
||||
cb(kv_cmpr, "kv_cmpr", il);
|
||||
|
||||
// MLA attention
|
||||
{
|
||||
// {n_embd_head_qk_nope, n_tokens, n_head}
|
||||
q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
|
||||
cb(q_nope, "q_nope_perm", il);
|
||||
|
||||
// {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
|
||||
ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
|
||||
cb(q_nope_absorbed, "q_nope_absorbed", il);
|
||||
|
||||
// {kv_lora_rank, n_head, n_tokens}
|
||||
q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
|
||||
cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
|
||||
|
||||
// {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
|
||||
// note: rope must go first for in-place context shifting in build_rope_shift()
|
||||
ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
|
||||
cb(Qcur, "Qcur", il);
|
||||
|
||||
kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
|
||||
cb(kv_cmpr, "kv_cmpr_reshape", il);
|
||||
|
||||
// {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
|
||||
ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
|
||||
cb(Kcur, "Kcur", il);
|
||||
|
||||
// {kv_lora_rank, 1, n_tokens}
|
||||
ggml_tensor * Vcur = kv_cmpr;
|
||||
cb(Vcur, "Vcur", il);
|
||||
|
||||
// note: MLA with the absorption optimization converts into MQA (ie: GQA with 1 group)
|
||||
cur = build_attn(inp_attn_dsa,
|
||||
model.layers[il].wo, NULL, model.layers[il].wo_s,
|
||||
Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, top_k, kq_scale, il);
|
||||
}
|
||||
}
|
||||
if (il == effective_n_layers - 1 && inp_out_ids) {
|
||||
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
|
||||
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
|
||||
}
|
||||
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
|
||||
cb(ffn_inp, "ffn_inp", il);
|
||||
|
||||
cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
|
||||
cb(cur, "ffn_norm", il);
|
||||
|
||||
if ((uint32_t) il < hparams.n_layer_dense_lead) {
|
||||
cur = build_ffn(cur,
|
||||
model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,
|
||||
model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
|
||||
model.layers[il].ffn_down, NULL, model.layers[il].ffn_down_s,
|
||||
NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
|
||||
cb(cur, "ffn_out", il);
|
||||
} else {
|
||||
// MoE branch
|
||||
ggml_tensor * moe_out = build_moe_ffn(cur,
|
||||
model.layers[il].ffn_gate_inp,
|
||||
model.layers[il].ffn_up_exps,
|
||||
model.layers[il].ffn_gate_exps,
|
||||
model.layers[il].ffn_down_exps,
|
||||
model.layers[il].ffn_exp_probs_b,
|
||||
n_expert, n_expert_used,
|
||||
LLM_FFN_SILU, hparams.expert_weights_norm,
|
||||
hparams.expert_weights_scale,
|
||||
(llama_expert_gating_func_type) hparams.expert_gating_func,
|
||||
il,
|
||||
nullptr,
|
||||
model.layers[il].ffn_gate_up_exps,
|
||||
model.layers[il].ffn_up_exps_s,
|
||||
model.layers[il].ffn_gate_exps_s,
|
||||
model.layers[il].ffn_down_exps_s);
|
||||
cb(moe_out, "ffn_moe_out", il);
|
||||
|
||||
// FFN shared expert
|
||||
{
|
||||
ggml_tensor * ffn_shexp =
|
||||
build_ffn(cur,
|
||||
model.layers[il].ffn_up_shexp, NULL, model.layers[il].ffn_up_shexp_s,
|
||||
model.layers[il].ffn_gate_shexp, NULL, model.layers[il].ffn_gate_shexp_s,
|
||||
model.layers[il].ffn_down_shexp, NULL, model.layers[il].ffn_down_shexp_s,
|
||||
NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
|
||||
cb(ffn_shexp, "ffn_shexp", il);
|
||||
|
||||
cur = ggml_add(ctx0, moe_out, ffn_shexp);
|
||||
cb(cur, "ffn_out", il);
|
||||
}
|
||||
}
|
||||
cur = ggml_add(ctx0, cur, ffn_inp);
|
||||
|
||||
cur = build_cvec(cur, il);
|
||||
cb(cur, "l_out", il);
|
||||
|
||||
// input for next layer
|
||||
inpL = cur;
|
||||
}
|
||||
cur = inpL;
|
||||
|
||||
cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
|
||||
|
||||
cb(cur, "result_norm", -1);
|
||||
res->t_embd = cur;
|
||||
|
||||
// lm_head
|
||||
cur = ggml_mul_mat(ctx0, model.output, cur);
|
||||
|
||||
cb(cur, "result_output", -1);
|
||||
res->t_logits = cur;
|
||||
|
||||
ggml_build_forward_expand(gf, cur);
|
||||
}
|
||||
@@ -1030,6 +1030,19 @@ struct llama_model_deepseek2 : public llama_model_base {
|
||||
};
|
||||
|
||||
|
||||
struct llama_model_deepseek32 : public llama_model_base {
|
||||
llama_model_deepseek32(const struct llama_model_params & params) : llama_model_base(params) {}
|
||||
void load_arch_hparams(llama_model_loader & ml) override;
|
||||
void load_arch_tensors(llama_model_loader & ml) override;
|
||||
|
||||
struct graph : public llm_graph_context {
|
||||
graph(const llama_model & model, const llm_graph_params & params);
|
||||
};
|
||||
|
||||
std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
|
||||
};
|
||||
|
||||
|
||||
struct llama_model_deepseek2ocr : public llama_model_base {
|
||||
llama_model_deepseek2ocr(const struct llama_model_params & params) : llama_model_base(params) {}
|
||||
void load_arch_hparams(llama_model_loader & ml) override;
|
||||
|
||||
+21
-8
@@ -508,28 +508,41 @@ llama_model_qwen35::graph_mtp::graph_mtp(const llama_model & model, const llm_gr
|
||||
int sections[4];
|
||||
std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
|
||||
|
||||
auto inp = std::make_unique<llm_graph_input_embd>(hparams.n_embd);
|
||||
// TODO: extract in a common llm_graph_context::build_inp_embd_h()
|
||||
auto inp = std::make_unique<llm_graph_input_embd_h>(hparams.n_embd);
|
||||
|
||||
inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
|
||||
ggml_set_input(inp->tokens);
|
||||
|
||||
inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
|
||||
inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens);
|
||||
ggml_set_input(inp->embd);
|
||||
ggml_set_name(inp->embd, "mtp_h_input");
|
||||
|
||||
ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;
|
||||
// TODO: make static using `ggml_build_forward_select()`
|
||||
// see llm_graph_context::build_inp_embd() for reference
|
||||
ggml_tensor * tok_embd;
|
||||
if (ubatch.token) {
|
||||
ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;
|
||||
|
||||
ggml_tensor * h_input = inp->embd;
|
||||
ggml_tensor * tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);
|
||||
tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);
|
||||
} else {
|
||||
tok_embd = inp->embd;
|
||||
}
|
||||
cb(tok_embd, "mtp_tok_embd", il);
|
||||
|
||||
inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
|
||||
ggml_set_input(inp->h);
|
||||
ggml_set_name(inp->h, "mtp_h_input");
|
||||
|
||||
ggml_tensor * h_embd = inp->h;
|
||||
|
||||
res->add_input(std::move(inp));
|
||||
|
||||
ggml_tensor * inp_pos = build_inp_pos();
|
||||
ggml_tensor * inp_out_ids = build_inp_out_ids();
|
||||
auto * inp_attn = build_attn_inp_kv();
|
||||
|
||||
ggml_tensor * h_norm = build_norm(h_input, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
|
||||
auto * inp_attn = build_attn_inp_kv();
|
||||
|
||||
ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
|
||||
cb(h_norm, "mtp_hnorm", il);
|
||||
|
||||
ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);
|
||||
|
||||
@@ -571,29 +571,41 @@ llama_model_qwen35moe::graph_mtp::graph_mtp(const llama_model & model, const llm
|
||||
int sections[4];
|
||||
std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
|
||||
|
||||
auto inp = std::make_unique<llm_graph_input_embd>(hparams.n_embd);
|
||||
// TODO: extract in a common llm_graph_context::build_inp_embd_h()
|
||||
auto inp = std::make_unique<llm_graph_input_embd_h>(hparams.n_embd);
|
||||
|
||||
inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
|
||||
ggml_set_input(inp->tokens);
|
||||
|
||||
inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
|
||||
inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens);
|
||||
ggml_set_input(inp->embd);
|
||||
ggml_set_name(inp->embd, "mtp_h_input");
|
||||
|
||||
ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;
|
||||
// TODO: make static using `ggml_build_forward_select()`
|
||||
// see llm_graph_context::build_inp_embd() for reference
|
||||
ggml_tensor * tok_embd;
|
||||
if (ubatch.token) {
|
||||
ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;
|
||||
|
||||
ggml_tensor * h_input = inp->embd;
|
||||
ggml_tensor * tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);
|
||||
tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);
|
||||
} else {
|
||||
tok_embd = inp->embd;
|
||||
}
|
||||
cb(tok_embd, "mtp_tok_embd", il);
|
||||
|
||||
inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
|
||||
ggml_set_input(inp->h);
|
||||
ggml_set_name(inp->h, "mtp_h_input");
|
||||
|
||||
ggml_tensor * h_embd = inp->h;
|
||||
|
||||
res->add_input(std::move(inp));
|
||||
|
||||
ggml_tensor * inp_pos = build_inp_pos();
|
||||
ggml_tensor * inp_out_ids = build_inp_out_ids();
|
||||
auto * inp_attn = build_attn_inp_kv();
|
||||
|
||||
auto * inp_attn = build_attn_inp_kv();
|
||||
|
||||
ggml_tensor * h_norm = build_norm(h_input, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
|
||||
ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
|
||||
cb(h_norm, "mtp_hnorm", il);
|
||||
|
||||
ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);
|
||||
|
||||
@@ -29,6 +29,7 @@ struct clip_graph {
|
||||
const int n_patches;
|
||||
const int n_embd;
|
||||
const int n_head;
|
||||
const int n_head_kv;
|
||||
const int d_head;
|
||||
const int n_layer;
|
||||
const int n_mmproj_embd;
|
||||
|
||||
@@ -188,6 +188,8 @@
|
||||
#define TN_SAM_FFN_DOWN "v.sam.blk.%d.mlp.lin2.%s"
|
||||
#define TN_SAM_NECK "v.sam.neck.%d.%s"
|
||||
#define TN_SAM_NET "v.sam.net_%d.%s"
|
||||
// deepseek-ocr-2
|
||||
#define TN_RESMPL_QUERY "v.resample_query_%d.%s"
|
||||
// (conformer) lfm2
|
||||
#define TN_PRE_ENCODE_OUT "a.pre_encode.out.%s"
|
||||
#define TN_FFN_NORM "%s.blk.%d.ffn_norm.%s"
|
||||
@@ -337,6 +339,7 @@ enum projector_type {
|
||||
PROJECTOR_TYPE_JANUS_PRO,
|
||||
PROJECTOR_TYPE_DOTS_OCR,
|
||||
PROJECTOR_TYPE_DEEPSEEKOCR,
|
||||
PROJECTOR_TYPE_DEEPSEEKOCR2,
|
||||
PROJECTOR_TYPE_LFM2A,
|
||||
PROJECTOR_TYPE_GLM4V,
|
||||
PROJECTOR_TYPE_YOUTUVL,
|
||||
@@ -386,6 +389,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
|
||||
{ PROJECTOR_TYPE_JANUS_PRO, "janus_pro"},
|
||||
{ PROJECTOR_TYPE_DOTS_OCR, "dots_ocr"},
|
||||
{ PROJECTOR_TYPE_DEEPSEEKOCR,"deepseekocr"},
|
||||
{ PROJECTOR_TYPE_DEEPSEEKOCR2,"deepseekocr2"},
|
||||
{ PROJECTOR_TYPE_LFM2A, "lfm2a"},
|
||||
{ PROJECTOR_TYPE_GLM4V, "glm4v"},
|
||||
{ PROJECTOR_TYPE_YOUTUVL, "youtuvl"},
|
||||
@@ -424,6 +428,9 @@ struct clip_image_f32 {
|
||||
int ny;
|
||||
|
||||
std::vector<float> buf;
|
||||
|
||||
// marks the global view in e.g., DeepSeek-OCR Models
|
||||
bool add_viewsep = false;
|
||||
};
|
||||
|
||||
//
|
||||
|
||||
@@ -542,6 +542,11 @@ struct clip_model {
|
||||
int32_t n_sam_layers = 12; // used by deepseek-ocr sam encoder
|
||||
|
||||
std::vector<clip_layer> sam_layers;
|
||||
|
||||
// deepseek-ocr-2
|
||||
ggml_tensor * resample_query_768 = nullptr;
|
||||
ggml_tensor * resample_query_1024 = nullptr;
|
||||
|
||||
// lfm2 audio
|
||||
std::array<ggml_tensor *, 7> pre_encode_conv_X_w = {nullptr};
|
||||
std::array<ggml_tensor *, 7> pre_encode_conv_X_b = {nullptr};
|
||||
|
||||
+69
-5
@@ -76,6 +76,7 @@
|
||||
#include "models/siglip.cpp"
|
||||
#include "models/whisper-enc.cpp"
|
||||
#include "models/deepseekocr.cpp"
|
||||
#include "models/deepseekocr2.cpp"
|
||||
#include "models/mobilenetv5.cpp"
|
||||
#include "models/youtuvl.cpp"
|
||||
#include "models/yasa2.cpp"
|
||||
@@ -306,6 +307,7 @@ clip_graph::clip_graph(clip_ctx * ctx, const clip_image_f32 & img) :
|
||||
n_patches(n_patches_x * n_patches_y),
|
||||
n_embd(hparams.n_embd),
|
||||
n_head(hparams.n_head),
|
||||
n_head_kv(hparams.n_head_kv),
|
||||
d_head(n_embd / n_head),
|
||||
n_layer(hparams.n_layer),
|
||||
n_mmproj_embd(clip_n_mmproj_embd(ctx)),
|
||||
@@ -461,9 +463,9 @@ ggml_tensor * clip_graph::build_vit(
|
||||
}
|
||||
}
|
||||
|
||||
Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_pos);
|
||||
Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_pos);
|
||||
Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_pos);
|
||||
Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_pos);
|
||||
Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head_kv, n_pos);
|
||||
Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head_kv, n_pos);
|
||||
|
||||
if (norm_per_head) {
|
||||
if (layer.q_norm) {
|
||||
@@ -1012,6 +1014,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
|
||||
{
|
||||
builder = std::make_unique<clip_graph_deepseekocr>(ctx, img);
|
||||
} break;
|
||||
case PROJECTOR_TYPE_DEEPSEEKOCR2:
|
||||
{
|
||||
builder = std::make_unique<clip_graph_deepseekocr2>(ctx, img);
|
||||
} break;
|
||||
case PROJECTOR_TYPE_LFM2A:
|
||||
{
|
||||
builder = std::make_unique<clip_graph_conformer>(ctx, img);
|
||||
@@ -1198,6 +1204,9 @@ struct clip_model_loader {
|
||||
get_u32(string_format(KEY_PROJ_DIM, prefix), hparams.projection_dim);
|
||||
get_f32(string_format(KEY_LAYER_NORM_EPS, prefix), hparams.eps);
|
||||
|
||||
// n_head_kv is optional (for GQA), default to n_head
|
||||
hparams.n_head_kv = hparams.n_head;
|
||||
|
||||
if (is_vision) {
|
||||
get_u32(KEY_IMAGE_SIZE, hparams.image_size);
|
||||
get_u32(KEY_PATCH_SIZE, hparams.patch_size);
|
||||
@@ -1598,6 +1607,7 @@ struct clip_model_loader {
|
||||
hparams.set_warmup_n_tokens(28*28); // avoid OOM on warmup
|
||||
} break;
|
||||
case PROJECTOR_TYPE_DEEPSEEKOCR:
|
||||
case PROJECTOR_TYPE_DEEPSEEKOCR2:
|
||||
{
|
||||
hparams.patch_size = 16;
|
||||
hparams.image_size = 1024;
|
||||
@@ -1609,6 +1619,10 @@ struct clip_model_loader {
|
||||
get_u32(KEY_SAM_N_HEAD, hparams.sam_n_head, true);
|
||||
get_u32(KEY_SAM_N_EMBD, hparams.sam_n_embd, true);
|
||||
get_u32(KEY_ATTN_WINDOW_SIZE, hparams.attn_window_size, true);
|
||||
if (model.proj_type == PROJECTOR_TYPE_DEEPSEEKOCR2) {
|
||||
// qwen2 encoder is GQA, requires KEY_N_HEAD_KV
|
||||
get_u32(string_format(KEY_N_HEAD_KV, "vision"), hparams.n_head_kv);
|
||||
}
|
||||
} break;
|
||||
case PROJECTOR_TYPE_HUNYUANVL:
|
||||
{
|
||||
@@ -1640,6 +1654,9 @@ struct clip_model_loader {
|
||||
hparams.audio_n_fft = 512;
|
||||
hparams.audio_window_len = 320; // 20ms frame (NOT 25ms/400)
|
||||
hparams.audio_hop_len = 160;
|
||||
// due to a mistake in the original conversion code, rms_norm_eps is set to a wrong value
|
||||
// since all gemma4a models use 1e-6, we just hardcode it here to avoid re-conversion
|
||||
hparams.eps = 1e-6f;
|
||||
} break;
|
||||
case PROJECTOR_TYPE_GRANITE_SPEECH:
|
||||
{
|
||||
@@ -2460,6 +2477,7 @@ struct clip_model_loader {
|
||||
model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
|
||||
} break;
|
||||
case PROJECTOR_TYPE_DEEPSEEKOCR:
|
||||
case PROJECTOR_TYPE_DEEPSEEKOCR2:
|
||||
{
|
||||
model.pos_embed = get_tensor(string_format(TN_SAM_POS_EMBD, "weight"));
|
||||
model.patch_embed_proj_w = get_tensor(string_format(TN_SAM_PATCH_EMBD, "weight"));
|
||||
@@ -2490,10 +2508,12 @@ struct clip_model_loader {
|
||||
model.neck_3_w = get_tensor(string_format(TN_SAM_NECK, 3, "weight"));
|
||||
model.net_2 = get_tensor(string_format(TN_SAM_NET, 2, "weight"));
|
||||
model.net_3 = get_tensor(string_format(TN_SAM_NET, 3, "weight"));
|
||||
model.image_newline = get_tensor(TN_IMAGE_NEWLINE);
|
||||
model.image_newline = get_tensor(TN_IMAGE_NEWLINE, false);
|
||||
model.view_seperator = get_tensor(TN_IMAGE_SEPERATOR);
|
||||
model.mm_fc_w = get_tensor(string_format(TN_MM_PROJECTOR, "weight"));
|
||||
model.mm_fc_b = get_tensor(string_format(TN_MM_PROJECTOR, "bias"));
|
||||
model.resample_query_768 = get_tensor(string_format(TN_RESMPL_QUERY, 768, "weight"), false);
|
||||
model.resample_query_1024 = get_tensor(string_format(TN_RESMPL_QUERY, 1024, "weight"), false);
|
||||
} break;
|
||||
case PROJECTOR_TYPE_GEMMA4A:
|
||||
{
|
||||
@@ -3493,6 +3513,11 @@ void setup_init_vision_shim_kcpp(struct clip_ctx * ctx_v) {
|
||||
img_end = "\n"; // prevent empty batch on llama-server
|
||||
image_preproc = std::make_unique<mtmd_image_preprocessor_deepseekocr>(ctx_v);
|
||||
} break;
|
||||
case PROJECTOR_TYPE_DEEPSEEKOCR2:
|
||||
{
|
||||
img_end = "\n"; // prevent empty batch on llama-server
|
||||
image_preproc = std::make_unique<mtmd_image_preprocessor_deepseekocr2>(ctx_v);
|
||||
} break;
|
||||
case PROJECTOR_TYPE_HUNYUANVL:
|
||||
{
|
||||
// note: these use fullwidth ｜ (U+FF5C) and ▁ (U+2581) to match the tokenizer vocabulary
|
||||
@@ -3789,7 +3814,7 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
|
||||
case PROJECTOR_TYPE_DEEPSEEKOCR:
|
||||
{
|
||||
// SAM encoder applies two stride-2 convolutions (net_2 and net_3)
|
||||
// which reduces spatial dimensions by 4x in each direction (16x total)
|
||||
// that reduce spatial dimensions by 4x in each direction (16x total)
|
||||
// E.g., 64x64 -> 16x16 patches
|
||||
n_patches /= 16;
|
||||
|
||||
@@ -3805,6 +3830,15 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
|
||||
int oh = (img->ny / patch_size) / merge;
|
||||
n_patches = (ow + 1) * oh + 2;
|
||||
} break;
|
||||
case PROJECTOR_TYPE_DEEPSEEKOCR2:
|
||||
{
|
||||
// 1024 global view -> 256 query tokens + 1 view separator = 257;
|
||||
// 768 local tile -> 144 query tokens, no separator.
|
||||
n_patches /= 16;
|
||||
if (img->add_viewsep) {
|
||||
n_patches += 1; // view separator, appended only after the global view
|
||||
}
|
||||
} break;
|
||||
case PROJECTOR_TYPE_LFM2A:
|
||||
{
|
||||
n_patches = ((((img->nx + 1) / 2) + 1) / 2 + 1) / 2;
|
||||
@@ -4394,6 +4428,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
|
||||
set_input_i32("pos_y", pos_y);
|
||||
} break;
|
||||
case PROJECTOR_TYPE_DEEPSEEKOCR:
|
||||
case PROJECTOR_TYPE_DEEPSEEKOCR2:
|
||||
{
|
||||
GGML_ASSERT(pos_w == pos_h);
|
||||
|
||||
@@ -4416,6 +4451,34 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
|
||||
|
||||
set_input_i32("rel_pos_indices_local", rel_pos_indices_local);
|
||||
set_input_i32("rel_pos_indices_global", rel_pos_indices_global);
|
||||
|
||||
if (ctx->proj_type() == PROJECTOR_TYPE_DEEPSEEKOCR2) {
|
||||
|
||||
// qwen2 encoder attention mask
|
||||
|
||||
// num_image_tokens = num_patches / 16
|
||||
// 256 for 1024 global view
|
||||
// 144 for 768 tile views
|
||||
const int num_image_tokens = num_patches / 16;
|
||||
const int seq_len = num_image_tokens * 2;
|
||||
std::vector qwen2_mask(static_cast<size_t>(seq_len) * seq_len, 0.0f);
|
||||
|
||||
// attention mask layout
|
||||
// +--------------+---------------+
|
||||
// | all 0 | all -inf |
|
||||
// +--------------+---------------+
|
||||
// | all 0 | lower tri 0 |
|
||||
// +--------------+---------------+
|
||||
for (int i = 0; i < seq_len; i++) {
|
||||
for (int j = 0; j < seq_len; j++) {
|
||||
const bool zero = i < num_image_tokens ?
|
||||
j < num_image_tokens :
|
||||
j < num_image_tokens || j <= i;
|
||||
qwen2_mask[static_cast<size_t>(i) * seq_len + j] = zero ? 0.0f : -1e9f;
|
||||
}
|
||||
}
|
||||
set_input_f32("qwen2_attn_mask", qwen2_mask);
|
||||
}
|
||||
} break;
|
||||
case PROJECTOR_TYPE_GEMMA3:
|
||||
case PROJECTOR_TYPE_GEMMA3NV:
|
||||
@@ -4965,6 +5028,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
|
||||
case PROJECTOR_TYPE_COGVLM:
|
||||
return ctx->model.mm_4h_to_h_w->ne[1];
|
||||
case PROJECTOR_TYPE_DEEPSEEKOCR:
|
||||
case PROJECTOR_TYPE_DEEPSEEKOCR2:
|
||||
return ctx->model.mm_fc_w->ne[1];
|
||||
case PROJECTOR_TYPE_LFM2A:
|
||||
return ctx->model.position_embeddings->ne[0];
|
||||
|
||||
@@ -30,7 +30,9 @@ static void show_additional_info(int /*argc*/, char ** argv) {
|
||||
" -p \"encode\" (debugging encode pass, default case):\n"
|
||||
" --image can be:\n"
|
||||
" \"white\", \"black\", \"gray\": filled 1.0f, 0.0f and 0.5f respectively\n"
|
||||
" \"red\", \"green\", \"blue\": filled with respective colors\n"
|
||||
" \"cb\": checkerboard pattern, alternate 1.0f and 0.0f\n"
|
||||
" \"rainbow\": raspberry-pi-like rainbow pattern\n"
|
||||
" --audio can be:\n"
|
||||
" \"one\", \"zero\", \"half\": filled 1.0f, 0.0f and 0.5f respectively\n"
|
||||
" \"1010\": checkerboard pattern, alternate 1.0f and 0.0f\n"
|
||||
@@ -144,6 +146,65 @@ int main(int argc, char ** argv) {
|
||||
image[y][x * 3 + 2] = v;
|
||||
}
|
||||
}
|
||||
} else if (input == "red") {
|
||||
for (int i = 0; i < inp_size; ++i) {
|
||||
auto row = std::vector<float>(inp_size * 3, 0.0f);
|
||||
for (int j = 0; j < inp_size; ++j) {
|
||||
row[j * 3 + 0] = 1.0f; // R channel
|
||||
}
|
||||
image.push_back(row);
|
||||
}
|
||||
} else if (input == "green") {
|
||||
for (int i = 0; i < inp_size; ++i) {
|
||||
auto row = std::vector<float>(inp_size * 3, 0.0f);
|
||||
for (int j = 0; j < inp_size; ++j) {
|
||||
row[j * 3 + 1] = 1.0f; // G channel
|
||||
}
|
||||
image.push_back(row);
|
||||
}
|
||||
} else if (input == "blue") {
|
||||
for (int i = 0; i < inp_size; ++i) {
|
||||
auto row = std::vector<float>(inp_size * 3, 0.0f);
|
||||
for (int j = 0; j < inp_size; ++j) {
|
||||
row[j * 3 + 2] = 1.0f; // B channel
|
||||
}
|
||||
image.push_back(row);
|
||||
}
|
||||
} else if (input == "rainbow") {
|
||||
for (int i = 0; i < inp_size; ++i) {
|
||||
image.push_back(std::vector<float>(inp_size * 3, 0.0f));
|
||||
}
|
||||
float cx = inp_size / 2.0f;
|
||||
float cy = inp_size / 2.0f;
|
||||
float max_dist = std::sqrt(cx * cx + cy * cy);
|
||||
for (int y = 0; y < inp_size; ++y) {
|
||||
for (int x = 0; x < inp_size; ++x) {
|
||||
float dx = x - cx;
|
||||
float dy = y - cy;
|
||||
float hue = std::atan2(dy, dx) / (2.0f * 3.14159265f);
|
||||
if (hue < 0) hue += 1.0f;
|
||||
float sat = std::sqrt(dx * dx + dy * dy) / max_dist;
|
||||
if (sat > 1.0f) sat = 1.0f;
|
||||
float h6 = hue * 6.0f;
|
||||
int i6 = (int)h6;
|
||||
float f = h6 - i6;
|
||||
float p = 1.0f - sat;
|
||||
float q = 1.0f - sat * f;
|
||||
float t = 1.0f - sat * (1.0f - f);
|
||||
float r, g, b;
|
||||
switch (i6 % 6) {
|
||||
case 0: r=1; g=t; b=p; break;
|
||||
case 1: r=q; g=1; b=p; break;
|
||||
case 2: r=p; g=1; b=t; break;
|
||||
case 3: r=p; g=q; b=1; break;
|
||||
case 4: r=t; g=p; b=1; break;
|
||||
default: r=1; g=p; b=q; break;
|
||||
}
|
||||
image[y][x * 3 + 0] = r;
|
||||
image[y][x * 3 + 1] = g;
|
||||
image[y][x * 3 + 2] = b;
|
||||
}
|
||||
}
|
||||
} else if (input == "one") {
|
||||
samples = std::vector<float>(inp_size, 1.0f);
|
||||
} else if (input == "zero") {
|
||||
|
||||
@@ -20,6 +20,43 @@ def test_vision():
|
||||
test_vision()
|
||||
```
|
||||
|
||||
Example of debugging a rainbow image:
|
||||
|
||||
```py
|
||||
import torch
|
||||
import math
|
||||
|
||||
def make_rainbow(img_size):
|
||||
cx, cy = img_size / 2.0, img_size / 2.0
|
||||
max_dist = math.sqrt(cx * cx + cy * cy)
|
||||
img = torch.zeros(1, 3, img_size, img_size)
|
||||
for y in range(img_size):
|
||||
for x in range(img_size):
|
||||
dx, dy = x - cx, y - cy
|
||||
hue = math.atan2(dy, dx) / (2 * math.pi)
|
||||
if hue < 0:
|
||||
hue += 1
|
||||
sat = math.sqrt(dx * dx + dy * dy) / max_dist
|
||||
sat = min(sat, 1.0)
|
||||
h6 = hue * 6
|
||||
i6 = int(h6)
|
||||
f = h6 - i6
|
||||
p = 1 - sat
|
||||
q = 1 - sat * f
|
||||
t = 1 - sat * (1 - f)
|
||||
rgb = [(1,t,p),(q,1,p),(p,1,t),(p,q,1),(t,p,1),(1,p,q)][i6 % 6]
|
||||
img[0, 0, y, x] = rgb[0]
|
||||
img[0, 1, y, x] = rgb[1]
|
||||
img[0, 2, y, x] = rgb[2]
|
||||
return img
|
||||
|
||||
img_size = 896
|
||||
pixel_values = make_rainbow(img_size)
|
||||
with torch.no_grad():
|
||||
outputs = model.model.get_image_features(pixel_values=pixel_values)
|
||||
print("last_hidden_state:", outputs.last_hidden_state)
|
||||
```
|
||||
|
||||
## Debugging preprocess pass
|
||||
|
||||
(TODO)
|
||||
|
||||
@@ -157,7 +157,6 @@ ggml_tensor * clip_graph_deepseekocr::build_sam(ggml_tensor * inp_raw) {
|
||||
|
||||
cur = ggml_mul_mat(ctx0, layer.qkv_w, cur);
|
||||
cur = ggml_add(ctx0, cur, layer.qkv_b);
|
||||
cur = ggml_cont(ctx0, cur); // Ensure tensor is contiguous before reshape
|
||||
cur = ggml_reshape_4d(ctx0, cur, n_embd, 3, W * H, B);
|
||||
|
||||
ggml_tensor * Q;
|
||||
@@ -251,17 +250,17 @@ ggml_cgraph * clip_graph_deepseekocr::build() {
|
||||
ggml_tensor * inp_raw = build_inp_raw();
|
||||
ggml_tensor * sam_out = build_sam(inp_raw);
|
||||
|
||||
const int clip_n_patches = sam_out->ne[0] * sam_out->ne[1];
|
||||
|
||||
ggml_tensor * clip_out;
|
||||
// Building DS-OCR CLIP
|
||||
{
|
||||
ggml_tensor * inp;
|
||||
|
||||
inp = ggml_cpy(ctx0, sam_out, ggml_dup_tensor(ctx0, sam_out));
|
||||
inp = ggml_reshape_2d(ctx0, inp, inp->ne[0] * inp->ne[1], inp->ne[2]);
|
||||
inp = ggml_reshape_2d(ctx0, sam_out, clip_n_patches, sam_out->ne[2]);
|
||||
inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));
|
||||
|
||||
ggml_tensor * new_pos_embd =
|
||||
ggml_cpy(ctx0, model.position_embeddings, ggml_dup_tensor(ctx0, model.position_embeddings));
|
||||
ggml_tensor * new_pos_embd = model.position_embeddings;
|
||||
|
||||
int n_pos = new_pos_embd->ne[1]; // +1 for [CLS]
|
||||
const auto tgt_size = static_cast<int>(std::sqrt(inp->ne[1]));
|
||||
@@ -295,16 +294,12 @@ ggml_cgraph * clip_graph_deepseekocr::build() {
|
||||
clip_out = cur;
|
||||
}
|
||||
|
||||
const int clip_n_patches = sam_out->ne[0] * sam_out->ne[1];
|
||||
|
||||
sam_out = ggml_cont(ctx0, ggml_permute(ctx0, sam_out, 1, 2, 0, 3));
|
||||
sam_out = ggml_reshape_2d(ctx0, sam_out, sam_out->ne[0], clip_n_patches);
|
||||
clip_out = ggml_view_2d(ctx0, clip_out, n_embd, clip_n_patches, clip_out->nb[1], clip_out->nb[1]);
|
||||
|
||||
ggml_tensor * cur;
|
||||
cur = ggml_concat(ctx0, clip_out, sam_out, 0);
|
||||
cur = ggml_reshape_2d(ctx0, cur, 2 * n_embd, clip_n_patches);
|
||||
cur = ggml_cont(ctx0, cur);
|
||||
cur = ggml_mul_mat(ctx0, model.mm_fc_w, cur);
|
||||
cur = ggml_add(ctx0, cur, model.mm_fc_b);
|
||||
|
||||
@@ -313,13 +308,11 @@ ggml_cgraph * clip_graph_deepseekocr::build() {
|
||||
const auto n_dim = cur->ne[0];
|
||||
|
||||
ggml_tensor * imgnl;
|
||||
ggml_tensor * vs;
|
||||
|
||||
imgnl = ggml_repeat_4d(ctx0, model.image_newline, n_dim, 1, h, 1);
|
||||
vs = ggml_reshape_2d(ctx0, model.view_seperator, n_dim, 1); // (n_dim, 1)
|
||||
cur = ggml_reshape_3d(ctx0, cur, n_dim, w, h);
|
||||
cur = ggml_reshape_2d(ctx0, ggml_concat(ctx0, cur, imgnl, 1), n_dim, (w + 1) * h);
|
||||
cur = ggml_concat(ctx0, cur, vs, 1); // (n_dim, h*(w+1) + 1)
|
||||
cur = ggml_concat(ctx0, cur, model.view_seperator, 1); // (n_dim, h*(w+1) + 1)
|
||||
|
||||
cb(cur, "dsocr_output", -1);
|
||||
|
||||
|
||||
@@ -0,0 +1,81 @@
|
||||
#include "models.h"
|
||||
|
||||
ggml_cgraph * clip_graph_deepseekocr2::build() {
|
||||
GGML_ASSERT(hparams.n_head_kv > 0);
|
||||
GGML_ASSERT(n_head % hparams.n_head_kv == 0);
|
||||
|
||||
// patch embedding
|
||||
ggml_tensor * inp_raw = build_inp_raw();
|
||||
|
||||
ggml_tensor * sam_out = build_sam(inp_raw);
|
||||
|
||||
ggml_tensor * qwen2_out;
|
||||
// Building Qwen2 encoder
|
||||
{
|
||||
ggml_tensor * inp;
|
||||
|
||||
inp = ggml_reshape_2d(ctx0, sam_out, sam_out->ne[0] * sam_out->ne[1], sam_out->ne[2]); // H*W, C
|
||||
inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));
|
||||
|
||||
auto num_image_tokens = inp->ne[1]; // H*W
|
||||
GGML_ASSERT(num_image_tokens == 144 || num_image_tokens == 256);
|
||||
|
||||
// query based on numbers of image tokens (in SAM output)
|
||||
// 16x16 -> query_1024 (1024x1024 images)
|
||||
// 12x12 -> query_768 (768x768 images)
|
||||
|
||||
ggml_tensor * query_embed = model.resample_query_1024;
|
||||
int num_queries = 256;
|
||||
|
||||
if (num_image_tokens == 144) {
|
||||
query_embed = model.resample_query_768;
|
||||
num_queries = 144;
|
||||
}
|
||||
|
||||
// (B, num_image_tokens + num_queries, C)
|
||||
inp = ggml_concat(ctx0, inp, ggml_cast(ctx0, query_embed, inp->type), 1);
|
||||
|
||||
auto seq_len = inp->ne[1];
|
||||
|
||||
// qwen2 encoder attention mask
|
||||
ggml_tensor * attn_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, seq_len, seq_len);
|
||||
ggml_set_name(attn_mask, "qwen2_attn_mask");
|
||||
ggml_set_input(attn_mask);
|
||||
|
||||
ggml_tensor * inp_pos = ggml_cast(ctx0, ggml_arange(ctx0, 0, seq_len, 1), GGML_TYPE_I32);
|
||||
|
||||
auto add_rope = [&](ggml_tensor * x, const clip_layer &) {
|
||||
return ggml_rope_ext(ctx0, x, inp_pos, nullptr, d_head,
|
||||
GGML_ROPE_TYPE_NEOX, 131072, 1000000, 1, 0, 1, 0, 0);
|
||||
};
|
||||
|
||||
build_vit_opts vit_opts;
|
||||
vit_opts.attn_mask = attn_mask;
|
||||
|
||||
// build_vit applies model.post_ln_w internally; do not re-apply
|
||||
ggml_tensor * cur = build_vit(inp, seq_len, NORM_TYPE_RMS, FFN_SILU,
|
||||
/* learned_pos_embd */ nullptr, add_rope, vit_opts);
|
||||
|
||||
cur = ggml_cont(ctx0,
|
||||
ggml_view_2d(ctx0, cur, cur->ne[0], num_queries, cur->nb[1],
|
||||
cur->nb[1] * (cur->ne[1] - num_queries))); // only take query tokens for output
|
||||
|
||||
ggml_build_forward_expand(gf, cur);
|
||||
qwen2_out = cur;
|
||||
}
|
||||
|
||||
ggml_tensor * cur;
|
||||
|
||||
cur = ggml_mul_mat(ctx0, model.mm_fc_w, qwen2_out);
|
||||
cur = ggml_add(ctx0, cur, model.mm_fc_b);
|
||||
|
||||
// view_seperator only after the global view
|
||||
if (img.add_viewsep) {
|
||||
cur = ggml_concat(ctx0, cur, model.view_seperator, 1); // (n_dim, 257)
|
||||
}
|
||||
|
||||
cb(cur, "dsocr2_output", -1);
|
||||
|
||||
ggml_build_forward_expand(gf, cur);
|
||||
return gf;
|
||||
}
|
||||
@@ -124,12 +124,12 @@ ggml_cgraph * clip_graph_gemma4v::build() {
|
||||
}
|
||||
|
||||
// Gemma4MultimodalEmbedder
|
||||
cur = build_mm(model.mm_input_proj_w, cur);
|
||||
cb(cur, "projected", -1);
|
||||
|
||||
// embedding_post_projection_norm
|
||||
cur = ggml_rms_norm(ctx0, cur, hparams.eps);
|
||||
cb(cur, "projected_normed", -1);
|
||||
{
|
||||
// embedding_pre_projection_norm
|
||||
cur = ggml_rms_norm(ctx0, cur, hparams.eps);
|
||||
cur = build_mm(model.mm_input_proj_w, cur);
|
||||
cb(cur, "projected", -1);
|
||||
}
|
||||
|
||||
ggml_build_forward_expand(gf, cur);
|
||||
return gf;
|
||||
|
||||
@@ -121,6 +121,11 @@ struct clip_graph_deepseekocr : clip_graph {
|
||||
ggml_tensor * build_sam(ggml_tensor * inp); // build the SAM model
|
||||
};
|
||||
|
||||
struct clip_graph_deepseekocr2 : clip_graph_deepseekocr {
|
||||
clip_graph_deepseekocr2(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph_deepseekocr(ctx, img) {}
|
||||
ggml_cgraph * build() override; // reuses build_sam() from base
|
||||
};
|
||||
|
||||
struct clip_graph_conformer : clip_graph {
|
||||
clip_graph_conformer(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
|
||||
ggml_cgraph * build() override;
|
||||
|
||||
@@ -1137,6 +1137,105 @@ bool mtmd_image_preprocessor_deepseekocr::preprocess(const clip_image_u8 & img,
|
||||
return true;
|
||||
}
|
||||
|
||||
//
|
||||
// mtmd_image_preprocessor_deepseekocr2
|
||||
//
|
||||
|
||||
// candidate tile grids (cols, rows) with min_tiles <= cols*rows <= max_tiles
|
||||
// sorted by tile count
|
||||
std::vector<clip_image_size> mtmd_image_preprocessor_deepseekocr2::get_target_ratios() {
|
||||
std::vector<clip_image_size> ratios;
|
||||
for (int n = min_tiles; n <= max_tiles; n++) {
|
||||
for (int w = 1; w <= n; w++) {
|
||||
for (int h = 1; h <= n; h++) {
|
||||
if (w * h < min_tiles || w * h > max_tiles) {
|
||||
continue;
|
||||
}
|
||||
bool found = false;
|
||||
for (const auto & r : ratios) {
|
||||
if (r.width == w && r.height == h) {
|
||||
found = true;
|
||||
break;
|
||||
}
|
||||
}
|
||||
if (!found) {
|
||||
ratios.push_back({ w, h });
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
std::sort(ratios.begin(), ratios.end(), [](const clip_image_size & a, const clip_image_size & b) {
|
||||
return a.width * a.height < b.width * b.height;
|
||||
});
|
||||
return ratios;
|
||||
}
|
||||
|
||||
// pick the grid whose aspect ratio is closest to the image
|
||||
// on a tie, prefer the larger grid when the image fits
|
||||
clip_image_size mtmd_image_preprocessor_deepseekocr2::find_closest_aspect_ratio(
|
||||
float aspect_ratio,
|
||||
const std::vector<clip_image_size> & target_ratios,
|
||||
int width,
|
||||
int height) {
|
||||
float best_ratio_diff = std::numeric_limits<float>::max();
|
||||
clip_image_size best_ratio = { 1, 1 };
|
||||
const float area = static_cast<float>(width * height);
|
||||
|
||||
for (const auto & ratio : target_ratios) {
|
||||
const float target_aspect_ratio = static_cast<float>(ratio.width) / ratio.height;
|
||||
const float ratio_diff = std::abs(aspect_ratio - target_aspect_ratio);
|
||||
if (ratio_diff < best_ratio_diff) {
|
||||
best_ratio_diff = ratio_diff;
|
||||
best_ratio = ratio;
|
||||
} else if (ratio_diff == best_ratio_diff) {
|
||||
const float target_area = static_cast<float>(tile_size * tile_size * ratio.width * ratio.height);
|
||||
if (area > 0.5f * target_area) {
|
||||
best_ratio = ratio;
|
||||
}
|
||||
}
|
||||
}
|
||||
return best_ratio;
|
||||
}
|
||||
|
||||
bool mtmd_image_preprocessor_deepseekocr2::preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) {
|
||||
// emit 768x768 local tiles when the image is larger than a tile in either
|
||||
// dimension, then always a 1024x1024 global view. order: [tiles..., global].
|
||||
|
||||
if (img.nx > tile_size || img.ny > tile_size) {
|
||||
const float aspect_ratio = static_cast<float>(img.nx) / img.ny;
|
||||
const auto target_ratios = get_target_ratios();
|
||||
const clip_image_size grid = find_closest_aspect_ratio(aspect_ratio, target_ratios, img.nx, img.ny);
|
||||
|
||||
// stretch onto the grid (no aspect preserve), then crop tiles row-major.
|
||||
clip_image_u8 refined;
|
||||
img_tool::resize(img, refined, { tile_size * grid.width, tile_size * grid.height },
|
||||
RESIZE_ALGO_BICUBIC_PILLOW, PAD_NONE);
|
||||
|
||||
for (int row = 0; row < grid.height; row++) {
|
||||
for (int col = 0; col < grid.width; col++) {
|
||||
clip_image_u8 tile;
|
||||
img_tool::crop(refined, tile, col * tile_size, row * tile_size, tile_size, tile_size);
|
||||
clip_image_f32_ptr res(clip_image_f32_init());
|
||||
img_u8_to_f32(tile, *res, hparams.image_mean, hparams.image_std);
|
||||
output.entries.push_back(std::move(res));
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// global view: aspect-preserving fit-and-pad to base_size.
|
||||
clip_image_u8 padded;
|
||||
img_tool::resize(img, padded, { base_size, base_size }, RESIZE_ALGO_BICUBIC_PILLOW,
|
||||
PAD_NEAREST, hparams.image_pad_color);
|
||||
clip_image_f32_ptr global(clip_image_f32_init());
|
||||
img_u8_to_f32(padded, *global, hparams.image_mean, hparams.image_std);
|
||||
global->add_viewsep = true;
|
||||
output.entries.push_back(std::move(global));
|
||||
|
||||
output.grid_x = 1;
|
||||
output.grid_y = 1;
|
||||
return true;
|
||||
}
|
||||
|
||||
//
|
||||
// mtmd_image_preprocessor_step3vl
|
||||
//
|
||||
|
||||
@@ -144,6 +144,26 @@ struct mtmd_image_preprocessor_deepseekocr : mtmd_image_preprocessor {
|
||||
bool preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) override;
|
||||
};
|
||||
|
||||
// DeepSeek-OCR-2: a 1024x1024 global view, plus InternVL-style 768x768 local
|
||||
// tiles when the image is larger than a tile in either dimension.
|
||||
struct mtmd_image_preprocessor_deepseekocr2 : mtmd_image_preprocessor {
|
||||
static constexpr int base_size = 1024; // global view
|
||||
static constexpr int tile_size = 768; // local tile
|
||||
static constexpr int min_tiles = 2;
|
||||
static constexpr int max_tiles = 6;
|
||||
|
||||
mtmd_image_preprocessor_deepseekocr2(const clip_ctx * ctx) : mtmd_image_preprocessor(ctx) {}
|
||||
bool preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) override;
|
||||
|
||||
private:
|
||||
static std::vector<clip_image_size> get_target_ratios();
|
||||
static clip_image_size find_closest_aspect_ratio(
|
||||
float aspect_ratio,
|
||||
const std::vector<clip_image_size> & target_ratios,
|
||||
int width,
|
||||
int height);
|
||||
};
|
||||
|
||||
// custom image preprocessing for Step3VL
|
||||
// ref: https://huggingface.co/stepfun-ai/Step3-VL-10B/blob/main/processing_step3.py
|
||||
struct mtmd_image_preprocessor_step3vl : mtmd_image_preprocessor_llava_uhd {
|
||||
|
||||
+12
-2
@@ -493,6 +493,11 @@ struct mtmd_context {
|
||||
img_end = "\n"; // prevent empty batch on llama-server
|
||||
image_preproc = std::make_unique<mtmd_image_preprocessor_deepseekocr>(ctx_v);
|
||||
} break;
|
||||
case PROJECTOR_TYPE_DEEPSEEKOCR2:
|
||||
{
|
||||
img_end = "\n"; // prevent empty batch on llama-server
|
||||
image_preproc = std::make_unique<mtmd_image_preprocessor_deepseekocr2>(ctx_v);
|
||||
} break;
|
||||
case PROJECTOR_TYPE_HUNYUANVL:
|
||||
{
|
||||
// note: these use fullwidth ｜ (U+FF5C) and ▁ (U+2581) to match the tokenizer vocabulary
|
||||
@@ -1091,16 +1096,21 @@ int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens)
|
||||
if (clip_is_llava(ctx_clip)
|
||||
|| proj_type == PROJECTOR_TYPE_MINICPMV
|
||||
|| proj_type == PROJECTOR_TYPE_GLM_EDGE
|
||||
|| proj_type == PROJECTOR_TYPE_INTERNVL) {
|
||||
|| proj_type == PROJECTOR_TYPE_INTERNVL
|
||||
|| proj_type == PROJECTOR_TYPE_DEEPSEEKOCR2) {
|
||||
// TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode()
|
||||
const auto & entries = image_tokens->batch_f32.entries;
|
||||
// entries may have different token counts
|
||||
// e.g., DeepSeek-OCR-2: 144 per tile views, 257 for the global view
|
||||
size_t offset = 0;
|
||||
for (size_t i = 0; i < entries.size(); i++) {
|
||||
int n_tokens_per_image = clip_n_output_tokens(ctx_clip, entries[i].get());
|
||||
ok = clip_image_encode(
|
||||
ctx_clip,
|
||||
ctx->n_threads,
|
||||
entries[i].get(),
|
||||
ctx->image_embd_v.data() + i*n_mmproj_embd*n_tokens_per_image);
|
||||
ctx->image_embd_v.data() + offset);
|
||||
offset += static_cast<size_t>(n_mmproj_embd) * n_tokens_per_image;
|
||||
}
|
||||
} else {
|
||||
ok = clip_image_batch_encode(
|
||||
|
||||
@@ -180,7 +180,8 @@ void server_model_meta::update_caps() {
|
||||
"LLAMA_ARG_HF_REPO",
|
||||
"LLAMA_ARG_HF_REPO_FILE",
|
||||
});
|
||||
params.offline = true; // avoid any unwanted network call during capability detection
|
||||
params.offline = true;
|
||||
// params.skip_download = true; // TODO: ideally, we should validate the model here, but it takes too much time
|
||||
common_params_handle_models(params, LLAMA_EXAMPLE_SERVER);
|
||||
if (params.mmproj.path.empty()) {
|
||||
multimodal = { false, false };
|
||||
@@ -371,18 +372,19 @@ void server_models::load_models() {
|
||||
// FIRST LOAD: add all models, then unlock for autoloading
|
||||
for (const auto & [name, preset] : final_presets) {
|
||||
server_model_meta meta{
|
||||
/* preset */ preset,
|
||||
/* name */ name,
|
||||
/* aliases */ {},
|
||||
/* tags */ {},
|
||||
/* port */ 0,
|
||||
/* status */ SERVER_MODEL_STATUS_UNLOADED,
|
||||
/* last_used */ 0,
|
||||
/* args */ std::vector<std::string>(),
|
||||
/* loaded_info */ {},
|
||||
/* exit_code */ 0,
|
||||
/* stop_timeout */ DEFAULT_STOP_TIMEOUT,
|
||||
/* multimodal */ mtmd_caps{false, false},
|
||||
/* preset */ preset,
|
||||
/* name */ name,
|
||||
/* aliases */ {},
|
||||
/* tags */ {},
|
||||
/* port */ 0,
|
||||
/* status */ SERVER_MODEL_STATUS_UNLOADED,
|
||||
/* last_used */ 0,
|
||||
/* args */ std::vector<std::string>(),
|
||||
/* loaded_info */ {},
|
||||
/* exit_code */ 0,
|
||||
/* stop_timeout */ DEFAULT_STOP_TIMEOUT,
|
||||
/* multimodal */ mtmd_caps{false, false},
|
||||
/* need_download */ false,
|
||||
};
|
||||
add_model(std::move(meta));
|
||||
}
|
||||
@@ -524,18 +526,19 @@ void server_models::load_models() {
|
||||
for (const auto & [name, preset] : final_presets) {
|
||||
if (mapping.find(name) == mapping.end()) {
|
||||
server_model_meta meta{
|
||||
/* preset */ preset,
|
||||
/* name */ name,
|
||||
/* aliases */ {},
|
||||
/* tags */ {},
|
||||
/* port */ 0,
|
||||
/* status */ SERVER_MODEL_STATUS_UNLOADED,
|
||||
/* last_used */ 0,
|
||||
/* args */ std::vector<std::string>(),
|
||||
/* loaded_info */ {},
|
||||
/* exit_code */ 0,
|
||||
/* stop_timeout */ DEFAULT_STOP_TIMEOUT,
|
||||
/* multimodal */ mtmd_caps{false, false},
|
||||
/* preset */ preset,
|
||||
/* name */ name,
|
||||
/* aliases */ {},
|
||||
/* tags */ {},
|
||||
/* port */ 0,
|
||||
/* status */ SERVER_MODEL_STATUS_UNLOADED,
|
||||
/* last_used */ 0,
|
||||
/* args */ std::vector<std::string>(),
|
||||
/* loaded_info */ {},
|
||||
/* exit_code */ 0,
|
||||
/* stop_timeout */ DEFAULT_STOP_TIMEOUT,
|
||||
/* multimodal */ mtmd_caps{false, false},
|
||||
/* need_download */ false,
|
||||
};
|
||||
add_model(std::move(meta));
|
||||
newly_added.push_back(name);
|
||||
@@ -1263,14 +1266,15 @@ void server_models_routes::init_routes() {
|
||||
};
|
||||
|
||||
json model_info = json {
|
||||
{"id", meta.name},
|
||||
{"aliases", meta.aliases},
|
||||
{"tags", meta.tags},
|
||||
{"object", "model"}, // for OAI-compat
|
||||
{"owned_by", "llamacpp"}, // for OAI-compat
|
||||
{"created", t}, // for OAI-compat
|
||||
{"status", status},
|
||||
{"architecture", architecture},
|
||||
{"id", meta.name},
|
||||
{"aliases", meta.aliases},
|
||||
{"tags", meta.tags},
|
||||
{"object", "model"}, // for OAI-compat
|
||||
{"owned_by", "llamacpp"}, // for OAI-compat
|
||||
{"created", t}, // for OAI-compat
|
||||
{"status", status},
|
||||
{"architecture", architecture},
|
||||
{"need_download", meta.need_download},
|
||||
// TODO: add other fields, may require reading GGUF metadata
|
||||
};
|
||||
|
||||
|
||||
@@ -67,6 +67,7 @@ struct server_model_meta {
|
||||
int exit_code = 0; // exit code of the model instance process (only valid if status == FAILED)
|
||||
int stop_timeout = 0; // seconds to wait before force-killing the model instance during shutdown
|
||||
mtmd_caps multimodal; // multimodal capabilities
|
||||
bool need_download = false; // whether the model needs to be downloaded before loading
|
||||
|
||||
bool is_ready() const {
|
||||
return status == SERVER_MODEL_STATUS_LOADED;
|
||||
|
||||
@@ -381,8 +381,10 @@ server_task_result_ptr server_response_reader::next(const std::function<bool()>
|
||||
if (result == nullptr) {
|
||||
// timeout, check stop condition
|
||||
if (should_stop()) {
|
||||
SRV_WRN("%s", "stopping wait for next result due to should_stop condition (adjust the --timeout argument if needed)\n");
|
||||
SRV_WRN("%s", "ref: https://github.com/ggml-org/llama.cpp/pull/22907\n");
|
||||
const int64_t time_elapsed_ms = ggml_time_ms() - time_start_ms;
|
||||
if (time_elapsed_ms > 30000) {
|
||||
SRV_WRN("%s", "request cancelled after 30s, potentially a client-side timeout; please check your client's code\n");
|
||||
}
|
||||
return nullptr;
|
||||
}
|
||||
} else {
|
||||
|
||||
@@ -169,6 +169,8 @@ struct server_response_reader {
|
||||
bool cancelled = false;
|
||||
int polling_interval_seconds;
|
||||
|
||||
const int64_t time_start_ms = ggml_time_ms();
|
||||
|
||||
// tracking generation state and partial tool calls
|
||||
// only used by streaming completions
|
||||
std::vector<task_result_state> states;
|
||||
|
||||
+4
@@ -106,10 +106,14 @@
|
||||
});
|
||||
|
||||
$effect(() => {
|
||||
void modelPropsVersion;
|
||||
|
||||
hasAudioModality = activeModelId ? modelsStore.modelSupportsAudio(activeModelId) : false;
|
||||
});
|
||||
|
||||
$effect(() => {
|
||||
void modelPropsVersion;
|
||||
|
||||
hasVideoModality = activeModelId ? modelsStore.modelSupportsVideo(activeModelId) : false;
|
||||
});
|
||||
|
||||
|
||||
Reference in New Issue
Block a user