Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 01:05:09 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'upstream' into concedo_experimental

Browse Source 
# Conflicts:
#	.flake8
#	.github/workflows/bench.yml
#	.github/workflows/python-lint.yml
#	.pre-commit-config.yaml
#	Makefile
#	README.md
#	models/ggml-vocab-bert-bge.gguf.inp
#	models/ggml-vocab-bert-bge.gguf.out
#	models/ggml-vocab-deepseek-coder.gguf.inp
#	models/ggml-vocab-deepseek-coder.gguf.out
#	models/ggml-vocab-deepseek-llm.gguf.inp
#	models/ggml-vocab-deepseek-llm.gguf.out
#	models/ggml-vocab-falcon.gguf.inp
#	models/ggml-vocab-falcon.gguf.out
#	models/ggml-vocab-gpt-2.gguf.inp
#	models/ggml-vocab-gpt-2.gguf.out
#	models/ggml-vocab-llama-bpe.gguf.inp
#	models/ggml-vocab-llama-bpe.gguf.out
#	models/ggml-vocab-llama-spm.gguf.inp
#	models/ggml-vocab-llama-spm.gguf.out
#	models/ggml-vocab-mpt.gguf.inp
#	models/ggml-vocab-mpt.gguf.out
#	models/ggml-vocab-phi-3.gguf
#	models/ggml-vocab-phi-3.gguf.inp
#	models/ggml-vocab-phi-3.gguf.out
#	models/ggml-vocab-refact.gguf
#	models/ggml-vocab-starcoder.gguf.inp
#	models/ggml-vocab-starcoder.gguf.out
#	requirements/requirements-convert.txt
#	scripts/compare-llama-bench.py
#	scripts/run-with-preset.py
#	scripts/verify-checksum-models.py
#	tests/CMakeLists.txt
#	tests/test-tokenizer-0.cpp
This commit is contained in:
Concedo
2024-05-06 18:09:45 +08:00
parent 173c7272d5 bcdee0daa7
commit 6c000cbe7a
40 changed files with 1593 additions and 936 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/gguf-split/gguf-split.cpp
+15 -4
View File
@@ -33,6 +33,7 @@ struct split_params {
int n_split_tensors = 128;
std::string input;
std::string output;
bool no_tensor_first_split = false;
bool dry_run = false;
};
@@ -50,6 +51,7 @@ static void split_print_usage(const char * executable) {
printf(" --merge merge multiple GGUF to a single GGUF\n");
printf(" --split-max-tensors max tensors in each split (default: %d)\n", default_params.n_split_tensors);
printf(" --split-max-size N(M|G) max size per split\n");
printf(" --no-tensor-first-split do not add tensors to the first split (disabled by default)\n");
printf(" --dry-run only print out a split plan and exit, without writing any new files\n");
printf("\n");
}
@@ -101,6 +103,10 @@ static void split_params_parse_ex(int argc, const char ** argv, split_params & p
arg_found = true;
params.dry_run = true;
}
if (arg == "--no-tensor-first-split") {
arg_found = true;
params.no_tensor_first_split = true;
}
if (is_op_set) {
throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
@@ -201,10 +207,10 @@ struct split_strategy {
// because we need to know list of tensors for each file in advance, we will build all the ctx_out for all output splits
int i_split = -1;
struct gguf_context * ctx_out = NULL;
auto new_ctx_out = [&]() {
auto new_ctx_out = [&](bool allow_no_tensors) {
i_split++;
if (ctx_out != NULL) {
if (gguf_get_n_tensors(ctx_out) == 0) {
if (gguf_get_n_tensors(ctx_out) == 0 && !allow_no_tensors) {
fprintf(stderr, "error: one of splits have 0 tensors. Maybe size or tensors limit is too small\n");
exit(EXIT_FAILURE);
}
@@ -221,7 +227,12 @@ struct split_strategy {
};
// initialize ctx_out for the first split
new_ctx_out();
new_ctx_out(false);
// skip first split if no_tensor_first_split is set
if (params.no_tensor_first_split) {
new_ctx_out(true);
}
// process tensors one by one
size_t curr_tensors_size = 0; // current size by counting only tensors size (without metadata)
@@ -231,7 +242,7 @@ struct split_strategy {
size_t n_bytes = GGML_PAD(ggml_nbytes(t), GGUF_DEFAULT_ALIGNMENT);
size_t next_tensors_size = curr_tensors_size + n_bytes;
if (should_split(i, next_tensors_size)) {
new_ctx_out();
new_ctx_out(false);
curr_tensors_size = n_bytes;
} else {
curr_tensors_size = next_tensors_size;
examples/gguf-split/tests.sh
+7 -7
View File
@@ -55,15 +55,15 @@ $MAIN --model $WORK_PATH/ggml-model-merge.gguf --random-prompt --n-predict 32
echo PASS
echo
# 4. Split with no tensor in metadata
#$SPLIT --split-max-tensors 32 --no-tensor-in-metadata $WORK_PATH/ggml-model-merge.gguf $WORK_PATH/ggml-model-split-32-tensors
#echo PASS
#echo
# 4. Split with no tensors in the first split
$SPLIT --split-max-tensors 32 --no-tensor-first-split $WORK_PATH/ggml-model-merge.gguf $WORK_PATH/ggml-model-split-32-tensors
echo PASS
echo
# 4b. Test the sharded model is loading properly
#$MAIN --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00006.gguf --random-prompt --n-predict 32
#echo PASS
#echo
$MAIN --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00007.gguf --random-prompt --n-predict 32
echo PASS
echo
# 5. Merge
#$SPLIT --merge $WORK_PATH/ggml-model-split-32-tensors-00001-of-00006.gguf $WORK_PATH/ggml-model-merge-2.gguf
examples/llama-bench/llama-bench.cpp
+15
View File
@@ -179,6 +179,7 @@ struct cmd_params {
std::vector<std::vector<float>> tensor_split;
std::vector<bool> use_mmap;
std::vector<bool> embeddings;
ggml_numa_strategy numa;
int reps;
bool verbose;
output_formats output_format;
@@ -201,6 +202,7 @@ static const cmd_params cmd_params_defaults = {
/* tensor_split */ {std::vector<float>(llama_max_devices(), 0.0f)},
/* use_mmap */ {true},
/* embeddings */ {false},
/* numa */ GGML_NUMA_STRATEGY_DISABLED,
/* reps */ 5,
/* verbose */ false,
/* output_format */ MARKDOWN
@@ -225,6 +227,7 @@ static void print_usage(int /* argc */, char ** argv) {
printf(" -nkvo, --no-kv-offload <0|1> (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
printf(" -fa, --flash-attn <0|1> (default: %s)\n", join(cmd_params_defaults.flash_attn, ",").c_str());
printf(" -mmp, --mmap <0|1> (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str());
printf(" --numa <distribute|isolate|numactl> (default: disabled)\n");
printf(" -embd, --embeddings <0|1> (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
printf(" -ts, --tensor-split <ts0/ts1/..> (default: 0)\n");
printf(" -r, --repetitions <n> (default: %d)\n", cmd_params_defaults.reps);
@@ -397,6 +400,17 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
}
auto p = split<bool>(argv[i], split_delim);
params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end());
} else if (arg == "--numa") {
if (++i >= argc) {
invalid_param = true;
break;
} else {
std::string value(argv[i]);
/**/ if (value == "distribute" || value == "" ) { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; }
else { invalid_param = true; break; }
}
} else if (arg == "-fa" || arg == "--flash-attn") {
if (++i >= argc) {
invalid_param = true;
@@ -1216,6 +1230,7 @@ int main(int argc, char ** argv) {
llama_log_set(llama_null_log_callback, NULL);
}
llama_backend_init();
llama_numa_init(params.numa);
// initialize printer
std::unique_ptr<printer> p;
examples/main/main.cpp
+1 -1
View File
@@ -545,7 +545,7 @@ int main(int argc, char ** argv) {
// if we run out of context:
// - take the n_keep first tokens from the original prompt (via n_past)
// - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) >= n_ctx) {
if (params.n_predict == -2) {
LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
break;
examples/server/server.cpp
+2 -1
View File
@@ -1384,9 +1384,10 @@ struct server_context {
if (!slot.params.stream && slot.stopped_word) {
const std::vector<llama_token> stop_word_toks = llama_tokenize(ctx, slot.stopping_word, false);
size_t safe_offset = std::min(slot.generated_token_probs.size(), stop_word_toks.size());
probs = std::vector<completion_token_output>(
slot.generated_token_probs.begin(),
slot.generated_token_probs.end() - stop_word_toks.size());
slot.generated_token_probs.end() - safe_offset);
} else {
probs = std::vector<completion_token_output>(
slot.generated_token_probs.begin(),
examples/server/tests/features/results.feature
+57 -33
View File
@@ -7,44 +7,16 @@ Feature: Results
And a model file tinyllamas/split/stories15M-00001-of-00003.gguf from HF repo ggml-org/models
And a model file test-model-00001-of-00003.gguf
And 128 as batch size
And 256 KV cache size
And 1024 KV cache size
And 128 max tokens to predict
Scenario Outline: Multi users completion
Given <n_slots> slots
And continuous batching
Scenario Outline: consistent results with same seed
Given <n_slots> slots
Then the server is starting
Then the server is healthy
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 4 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 42
Given concurrent completion requests
Then the server is busy
@@ -55,3 +27,55 @@ Feature: Results
| n_slots |
| 1 |
| 2 |
Scenario Outline: different results with different seed
Given <n_slots> slots
Then the server is starting
Then the server is healthy
Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 42
Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 43
Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 44
Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 45
Given concurrent completion requests
Then the server is busy
Then the server is idle
And all slots are idle
Then all predictions are different
Examples:
| n_slots |
| 1 |
| 2 |
Scenario Outline: consistent results with same seed and varying batch size
Given 4 slots
And <temp> temperature
# And 0 as draft
Then the server is starting
Then the server is healthy
Given 1 prompts "Write a very long story about AI." with seed 42
And concurrent completion requests
# Then the server is busy # Not all slots will be utilized.
Then the server is idle
And all slots are idle
Given <n_parallel> prompts "Write a very long story about AI." with seed 42
And concurrent completion requests
# Then the server is busy # Not all slots will be utilized.
Then the server is idle
And all slots are idle
Then all predictions are equal
Examples:
| n_parallel | temp |
| 1 | 0.0 |
| 2 | 0.0 |
| 4 | 0.0 |
| 1 | 1.0 |
# FIXME: These tests fail on master. The problem seems to be the unified KV cache.
# See https://github.com/ggerganov/whisper.cpp/issues/1941#issuecomment-1986923227
# and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574 .
# | 2 | 1.0 |
# | 4 | 1.0 |
examples/server/tests/features/steps/steps.py
+98 -46
View File
@@ -65,6 +65,7 @@ def step_server_config(context, server_fqdn, server_port):
context.server_seed = None
context.user_api_key = None
context.response_format = None
context.temperature = None
context.tasks_result = []
context.concurrent_tasks = []
@@ -232,15 +233,17 @@ async def step_all_slots_status(context, expected_slot_status_string):
@async_run_until_complete
async def step_request_completion(context, api_error):
expect_api_error = api_error == 'raised'
seeds = await completions_seed(context, num_seeds=1)
completion = await request_completion(context.prompts.pop(),
seeds[0] if seeds is not None else seeds,
context.base_url,
debug=context.debug,
n_predict=context.n_predict,
cache_prompt=context.cache_prompt,
id_slot=context.id_slot,
seed=await completions_seed(context),
expect_api_error=expect_api_error,
user_api_key=context.user_api_key)
user_api_key=context.user_api_key,
temperature=context.temperature)
context.tasks_result.append(completion)
if context.debug:
print(f"Completion response: {completion}")
@@ -269,6 +272,15 @@ async def step_predictions_equal(context):
context.tasks_result = []
@step('all predictions are different')
@async_run_until_complete
async def step_predictions_equal(context):
n_completions = await gather_tasks_results(context)
assert n_completions >= 2, "need at least 2 completions"
assert_all_predictions_different(context.tasks_result)
context.tasks_result = []
@step('the completion is truncated')
def step_assert_completion_truncated(context):
step_assert_completion_truncated(context, '')
@@ -311,6 +323,11 @@ def step_response_format(context, response_format):
context.response_format = json.loads(response_format)
@step('{temperature:f} temperature')
def step_temperature(context, temperature):
context.temperature = temperature
@step('streaming is {enable_streaming}')
def step_streaming(context, enable_streaming):
context.enable_streaming = enable_streaming == 'enabled'
@@ -353,7 +370,10 @@ def step_n_ubatch(context, n_ubatch):
@step('{seed:d} as seed')
def step_seed(context, seed):
context.seed = seed
if context.seed is None:
context.seed = [seed]
else:
context.seed.append(seed)
@step('a prefix prompt')
@@ -413,7 +433,9 @@ async def step_oai_chat_completions(context, api_error):
if context.debug:
print(f"Submitting OAI compatible completions request...")
expect_api_error = api_error == 'raised'
seeds = await completions_seed(context, num_seeds=1),
completion = await oai_chat_completions(context.prompts.pop(),
seeds[0] if seeds is not None else seeds,
context.system_prompt,
context.base_url,
'/v1/chat',
@@ -429,8 +451,6 @@ async def step_oai_chat_completions(context, api_error):
response_format=context.response_format
if hasattr(context, 'response_format') else None,
seed=await completions_seed(context),
user_api_key=context.user_api_key
if hasattr(context, 'user_api_key') else None,
@@ -457,20 +477,31 @@ def step_a_prompt_prompt(context, prompt):
context.n_prompts = len(context.prompts)
@step('{num_prompts:d} prompts {prompt} with seed {seed:d}')
def step_many_prompts(context, num_prompts, prompt, seed):
if context.seed is None:
context.seed = []
for _ in range(num_prompts):
context.seed.append(seed)
context.prompts.append(prompt)
context.n_prompts = len(context.prompts)
@step('concurrent completion requests')
@async_run_until_complete()
async def step_concurrent_completion_requests(context):
await concurrent_requests(context,
request_completion,
# prompt is inserted automatically
context.base_url,
debug=context.debug,
prompt_prefix=context.prompt_prefix,
prompt_suffix=context.prompt_suffix,
n_predict=context.n_predict if hasattr(context, 'n_predict') else None,
seed=await completions_seed(context),
user_api_key=context.user_api_key if hasattr(context,
'user_api_key') else None)
await concurrent_requests(
context,
request_completion,
# prompt is inserted automatically
context.base_url,
debug=context.debug,
prompt_prefix=context.prompt_prefix,
prompt_suffix=context.prompt_suffix,
n_predict=context.n_predict if hasattr(context, 'n_predict') else None,
user_api_key=context.user_api_key if hasattr(context, 'user_api_key') else None,
temperature=context.temperature,
)
@step('concurrent OAI completions requests')
@@ -490,7 +521,6 @@ async def step_oai_chat_completions(context):
if hasattr(context, 'enable_streaming') else None,
response_format=context.response_format
if hasattr(context, 'response_format') else None,
seed=await completions_seed(context),
user_api_key=context.user_api_key
if hasattr(context, 'user_api_key') else None)
@@ -512,10 +542,6 @@ async def step_oai_chat_completions(context):
if hasattr(context, 'enable_streaming') else None,
response_format=context.response_format
if hasattr(context, 'response_format') else None,
seed=context.seed
if hasattr(context, 'seed') else
context.server_seed
if hasattr(context, 'server_seed') else None,
user_api_key=context.user_api_key
if hasattr(context, 'user_api_key') else None)
@@ -544,7 +570,7 @@ async def all_prompts_are_predicted(context, expected_predicted_n=None):
@async_run_until_complete
async def step_compute_embedding(context):
context.n_prompts = 1
context.embeddings = await request_embedding(context_text(context), base_url=context.base_url)
context.embeddings = await request_embedding(context_text(context), None, base_url=context.base_url)
@step('all embeddings are the same')
@@ -585,7 +611,7 @@ def step_assert_embeddings(context):
@async_run_until_complete
async def step_oai_compute_embeddings(context):
context.n_prompts = 1
context.embeddings = await request_oai_embeddings(context_text(context),
context.embeddings = await request_oai_embeddings(context_text(context), None,
base_url=context.base_url,
user_api_key=context.user_api_key,
model=context.model)
@@ -594,7 +620,7 @@ async def step_oai_compute_embeddings(context):
@step('an OAI compatible embeddings computation request for multiple inputs')
@async_run_until_complete
async def step_oai_compute_embeddings_multiple_inputs(context):
context.embeddings = await request_oai_embeddings(context.prompts,
context.embeddings = await request_oai_embeddings(context.prompts, None,
base_url=context.base_url,
user_api_key=context.user_api_key,
model=context.model)
@@ -740,8 +766,9 @@ async def concurrent_requests(context, f_completion, *args, **kwargs):
if context.debug:
print(f"starting {context.n_prompts} concurrent completion requests...")
assert context.n_prompts > 0
seeds = await completions_seed(context)
for prompt_no in range(context.n_prompts):
shifted_args = [context.prompts.pop(), *args]
shifted_args = [context.prompts.pop(), seeds[prompt_no], *args]
context.concurrent_tasks.append(asyncio.create_task(f_completion(*shifted_args, **kwargs)))
await asyncio.sleep(0.1)
@@ -781,6 +808,7 @@ def step_server_responds_with_status_code(context, status_code):
async def request_completion(prompt,
seed,
base_url,
debug=False,
prompt_prefix=None,
@@ -788,9 +816,9 @@ async def request_completion(prompt,
n_predict=None,
cache_prompt=False,
id_slot=None,
seed=None,
expect_api_error=None,
user_api_key=None):
user_api_key=None,
temperature=None):
if debug:
print(f"Sending completion request: {prompt}")
origin = "my.super.domain"
@@ -811,7 +839,8 @@ async def request_completion(prompt,
"n_predict": n_predict if n_predict is not None else -1,
"cache_prompt": cache_prompt,
"id_slot": id_slot,
"seed": seed if seed is not None else 42
"seed": seed if seed is not None else 42,
"temperature": temperature if temperature is not None else "0.8f",
},
headers=headers,
timeout=3600) as response:
@@ -824,6 +853,7 @@ async def request_completion(prompt,
async def oai_chat_completions(user_prompt,
seed,
system_prompt,
base_url,
base_path,
@@ -833,7 +863,6 @@ async def oai_chat_completions(user_prompt,
n_predict=None,
enable_streaming=None,
response_format=None,
seed=None,
user_api_key=None,
expect_api_error=None):
if debug:
@@ -952,7 +981,7 @@ async def oai_chat_completions(user_prompt,
return completion_response
async def request_embedding(content, base_url=None):
async def request_embedding(content, seed, base_url=None):
async with aiohttp.ClientSession() as session:
async with session.post(f'{base_url}/embedding',
json={
@@ -963,7 +992,7 @@ async def request_embedding(content, base_url=None):
return [response_json['embedding']]
async def request_oai_embeddings(input,
async def request_oai_embeddings(input, seed,
base_url=None, user_api_key=None,
model=None, async_client=False):
# openai client always expects an api_key
@@ -1036,21 +1065,31 @@ def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re
f' {n_predicted} <> {expected_predicted_n}')
def assert_all_predictions_equal(completion_responses):
content_0 = completion_responses[0]['content']
if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
print(f"content 0: {content_0}")
for i, response_i in enumerate(completion_responses):
content_i = response_i['content']
print(f"content {i}: {content_i}")
for i, response_i in enumerate(completion_responses):
content_i = response_i['content']
for j, response_j in enumerate(completion_responses):
if i == j:
continue
content_j = response_j['content']
assert content_i == content_j, "contents not equal"
i = 1
for response in completion_responses[1:]:
content = response['content']
if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
print(f"content {i}: {content}")
assert content == content_0, "contents not equal"
i += 1
def assert_all_predictions_different(completion_responses):
if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
for i, response_i in enumerate(completion_responses):
content_i = response_i['content']
print(f"content {i}: {content_i}")
for i, response_i in enumerate(completion_responses):
content_i = response_i['content']
for j, response_j in enumerate(completion_responses):
if i == j:
continue
content_j = response_j['content']
assert content_i != content_j, "contents not different"
async def gather_tasks_results(context):
@@ -1145,9 +1184,22 @@ def assert_slots_status(slots, expected_slots):
f" = {expected[key]} != {slot[key]}")
async def completions_seed(context):
return context.seed if hasattr(context, 'seed') and context.seed is not None \
else context.server_seed if hasattr(context, 'server_seed') else None
async def completions_seed(context, num_seeds=None):
if hasattr(context, "seed") and context.seed is not None:
assert len(context.seed) == context.n_prompts
if num_seeds is None:
num_seeds = context.n_prompts
assert num_seeds <= context.n_prompts
seeds = context.seed[:num_seeds]
context.seed = context.seed[num_seeds:] if num_seeds < context.n_prompts else None
return seeds
if hasattr(context, "server_seed") and context.server_seed is not None:
if num_seeds is None:
return [context.server_seed] * context.n_prompts
else:
return [context.server_seed] * num_seeds
return None
def context_text(context):