From ff55414c42522adbeaa1bd9c52c0e9db16942484 Mon Sep 17 00:00:00 2001 From: "Piotr Wilkin (ilintar)" Date: Fri, 28 Nov 2025 12:02:56 +0100 Subject: [PATCH 1/2] model : Qwen3 Next (#16095) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit * Qwen3 Next - cleaned up version * Whitespaces and stuff * Correct minor errors * Update src/llama-model.cpp Co-authored-by: Sigbjørn Skjæret * Misc. fixes. * Clean up code, add missing hybrid qualifier * Did someone transpose the SOLVE_TRI result matrix? Perhaps... * Whitespace * Proper tensors for cb calls * Use llama-graph.h vertical alignment * BROKEN: chunking * Set new tensors as inputs. * Proper chunk logic * It's the circle of life... * More shenanigans for n_seq > 1 * Nail in the coffin? * Fix Windows build * Eh, one fails on Windows, the other fails on Mac... just use general capture. * quant : cleanup * model : cleanup * qwen3 : cleanup * cont : cleanup * cont : cleanup * ggml : revert change * qwen3 : cleanup * cont : cleanup * Readd cmath * qwen3 : fix typo * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret * Usual suspects * fix my bad suggestion --------- Co-authored-by: Sigbjørn Skjæret Co-authored-by: Georgi Gerganov --- convert_hf_to_gguf.py | 30 + .../scripts/causal/run-converted-model.sh | 8 +- .../scripts/causal/run-org-model.py | 8 +- ggml/src/ggml-cpu/ops.cpp | 3 +- gguf-py/gguf/constants.py | 33 + gguf-py/gguf/tensor_mapping.py | 22 +- src/CMakeLists.txt | 1 + src/llama-arch.cpp | 35 + src/llama-arch.h | 2 + src/llama-context.cpp | 4 + src/llama-hparams.h | 2 +- src/llama-model.cpp | 100 +- src/llama-model.h | 4 + src/llama-quant.cpp | 18 +- src/models/models.h | 52 +- src/models/qwen3next.cpp | 1042 +++++++++++++++++ 16 files changed, 1345 insertions(+), 19 deletions(-) create mode 100644 src/models/qwen3next.cpp diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index daaf0bf49..866aa536f 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -4183,6 +4183,36 @@ class Qwen3MoeModel(Qwen2MoeModel): super().set_vocab() +@ModelBase.register("Qwen3NextForCausalLM") +class Qwen3NextModel(Qwen2MoeModel): + model_arch = gguf.MODEL_ARCH.QWEN3NEXT + + def set_gguf_parameters(self): + super().set_gguf_parameters() + self.gguf_writer.add_ssm_conv_kernel(self.hparams["linear_conv_kernel_dim"]) + self.gguf_writer.add_ssm_state_size(self.hparams["linear_key_head_dim"]) + self.gguf_writer.add_ssm_group_count(self.hparams["linear_num_key_heads"]) + self.gguf_writer.add_ssm_time_step_rank(self.hparams["linear_num_value_heads"]) + self.gguf_writer.add_ssm_inner_size(self.hparams["linear_value_head_dim"] * self.hparams["linear_num_value_heads"]) + if (rope_dim := self.hparams.get("head_dim")) is None: + rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"] + self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.25))) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + if name.startswith("mtp"): + return [] # ignore MTP layers for now + if name.endswith(".A_log"): + data_torch = -torch.exp(data_torch) + elif name.endswith(".dt_bias"): + name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias" + elif "conv1d" in name: + data_torch = data_torch.squeeze() + elif name.endswith("norm.weight") and not name.endswith("linear_attn.norm.weight"): + data_torch = data_torch + 1 + + yield from super().modify_tensors(data_torch, name, bid) + + @ModelBase.register("RND1") class RND1Model(Qwen2MoeModel): model_arch = gguf.MODEL_ARCH.RND1 diff --git a/examples/model-conversion/scripts/causal/run-converted-model.sh b/examples/model-conversion/scripts/causal/run-converted-model.sh index f5f567d4f..529e9987b 100755 --- a/examples/model-conversion/scripts/causal/run-converted-model.sh +++ b/examples/model-conversion/scripts/causal/run-converted-model.sh @@ -4,6 +4,11 @@ set -e # First try command line argument, then environment variable, then file CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}" +MODEL_TESTING_PROMPT="${2:-"$MODEL_TESTING_PROMPT"}" + +if [ -z "$MODEL_TESTING_PROMPT"]; then + MODEL_TESTING_PROMPT="Hello, my name is" +fi # Final check if we have a model path if [ -z "$CONVERTED_MODEL" ]; then @@ -14,7 +19,8 @@ if [ -z "$CONVERTED_MODEL" ]; then fi echo $CONVERTED_MODEL +echo $MODEL_TESTING_PROMPT cmake --build ../../build --target llama-logits -j8 -../../build/bin/llama-logits -m "$CONVERTED_MODEL" "Hello, my name is" +../../build/bin/llama-logits -m "$CONVERTED_MODEL" "$MODEL_TESTING_PROMPT" diff --git a/examples/model-conversion/scripts/causal/run-org-model.py b/examples/model-conversion/scripts/causal/run-org-model.py index 85529c612..7d2b80057 100755 --- a/examples/model-conversion/scripts/causal/run-org-model.py +++ b/examples/model-conversion/scripts/causal/run-org-model.py @@ -184,8 +184,12 @@ model_name = os.path.basename(model_path) # of using AutoModelForCausalLM. print(f"Model class: {model.__class__.__name__}") -prompt = "Hello, my name is" -input_ids = tokenizer(prompt, return_tensors="pt").input_ids +device = next(model.parameters()).device +if os.getenv("MODEL_TESTING_PROMPT"): + prompt = os.getenv("MODEL_TESTING_PROMPT") +else: + prompt = "Hello, my name is" +input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device) print(f"Input tokens: {input_ids}") print(f"Input text: {repr(prompt)}") diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp index d40569653..2745fc54e 100644 --- a/ggml/src/ggml-cpu/ops.cpp +++ b/ggml/src/ggml-cpu/ops.cpp @@ -9766,7 +9766,8 @@ static void ggml_compute_forward_solve_tri_f32(const struct ggml_compute_params } const float diag = A_batch[i00 * n + i00]; - GGML_ASSERT(diag != 0.0f && "Zero diagonal in triangular matrix"); + assert(diag != 0.0f && "Zero diagonal in triangular matrix"); + X_batch[i00 * k + i01] = (B_batch[i00 * k + i01] - sum) / diag; } } diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 6f5a742e0..266d19f9d 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -366,6 +366,7 @@ class MODEL_ARCH(IntEnum): QWEN2VL = auto() QWEN3 = auto() QWEN3MOE = auto() + QWEN3NEXT = auto() QWEN3VL = auto() QWEN3VLMOE = auto() PHI2 = auto() @@ -531,6 +532,7 @@ class MODEL_TENSOR(IntEnum): SSM_D = auto() SSM_NORM = auto() SSM_OUT = auto() + SSM_BETA_ALPHA = auto() # qwen3next TIME_MIX_W0 = auto() TIME_MIX_W1 = auto() TIME_MIX_W2 = auto() @@ -736,6 +738,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.QWEN2VL: "qwen2vl", MODEL_ARCH.QWEN3: "qwen3", MODEL_ARCH.QWEN3MOE: "qwen3moe", + MODEL_ARCH.QWEN3NEXT: "qwen3next", MODEL_ARCH.QWEN3VL: "qwen3vl", MODEL_ARCH.QWEN3VLMOE: "qwen3vlmoe", MODEL_ARCH.PHI2: "phi2", @@ -900,6 +903,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = { MODEL_TENSOR.SSM_D: "blk.{bid}.ssm_d", MODEL_TENSOR.SSM_NORM: "blk.{bid}.ssm_norm", MODEL_TENSOR.SSM_OUT: "blk.{bid}.ssm_out", + MODEL_TENSOR.SSM_BETA_ALPHA: "blk.{bid}.ssm_ba", MODEL_TENSOR.TIME_MIX_W0: "blk.{bid}.time_mix_w0", MODEL_TENSOR.TIME_MIX_W1: "blk.{bid}.time_mix_w1", MODEL_TENSOR.TIME_MIX_W2: "blk.{bid}.time_mix_w2", @@ -1569,6 +1573,35 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN_EXP, MODEL_TENSOR.FFN_UP_EXP, ], + MODEL_ARCH.QWEN3NEXT: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_POST_NORM, + MODEL_TENSOR.ATTN_GATE, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_INP_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.SSM_A, + MODEL_TENSOR.SSM_CONV1D, + MODEL_TENSOR.SSM_DT, + MODEL_TENSOR.SSM_NORM, + MODEL_TENSOR.SSM_IN, + MODEL_TENSOR.SSM_BETA_ALPHA, + MODEL_TENSOR.SSM_OUT + ], MODEL_ARCH.QWEN3VL: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 8c7ed10f2..a7b097397 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -672,10 +672,11 @@ class TensorNameMap: ), MODEL_TENSOR.SSM_IN: ( - "model.layers.{bid}.in_proj", # mamba-hf - "backbone.layers.{bid}.mixer.in_proj", # mamba - "model.layers.{bid}.mamba.in_proj", # jamba falcon-h1 granite-hybrid - "model.layers.layers.{bid}.mixer.in_proj", # plamo2 + "model.layers.{bid}.in_proj", # mamba-hf + "backbone.layers.{bid}.mixer.in_proj", # mamba + "model.layers.{bid}.mamba.in_proj", # jamba falcon-h1 granite-hybrid + "model.layers.layers.{bid}.mixer.in_proj", # plamo2 + "model.layers.{bid}.linear_attn.in_proj_qkvz", # qwen3next ), MODEL_TENSOR.SSM_CONV1D: ( @@ -683,6 +684,7 @@ class TensorNameMap: "backbone.layers.{bid}.mixer.conv1d", # mamba "model.layers.{bid}.mamba.conv1d", # jamba falcon-h1 granite-hybrid "model.layers.layers.{bid}.mixer.conv1d", # plamo2 + "model.layers.{bid}.linear_attn.conv1d", # qwen3next ), MODEL_TENSOR.SSM_X: ( @@ -697,6 +699,7 @@ class TensorNameMap: "backbone.layers.{bid}.mixer.dt_proj", # mamba "model.layers.{bid}.mamba.dt_proj", # jamba falcon-h1 granite-hybrid "model.layers.layers.{bid}.mixer.dt_proj", # plamo2 + "model.layers.{bid}.linear_attn.dt_proj", # qwen3next ), MODEL_TENSOR.SSM_DT_NORM: ( @@ -709,6 +712,7 @@ class TensorNameMap: "backbone.layers.{bid}.mixer.A_log", # mamba "model.layers.{bid}.mamba.A_log", # jamba falcon-h1 granite-hybrid "model.layers.layers.{bid}.mixer.A_log", # plamo2 + "model.layers.{bid}.linear_attn.A_log", # qwen3next ), MODEL_TENSOR.SSM_B_NORM: ( @@ -731,17 +735,23 @@ class TensorNameMap: ), MODEL_TENSOR.SSM_NORM: ( - "model.layers.{bid}.mamba.norm", # falcon-h1 granite-hybrid - "backbone.layers.{bid}.mixer.norm", # mamba2 + "model.layers.{bid}.mamba.norm", # falcon-h1 granite-hybrid + "model.layers.{bid}.linear_attn.norm", # qwen3next + "backbone.layers.{bid}.mixer.norm", # mamba2 ), MODEL_TENSOR.SSM_OUT: ( "model.layers.{bid}.out_proj", # mamba-hf "backbone.layers.{bid}.mixer.out_proj", # mamba "model.layers.{bid}.mamba.out_proj", # jamba falcon-h1 granite-hybrid + "model.layers.{bid}.linear_attn.out_proj", # qwen3next "model.layers.layers.{bid}.mixer.out_proj", # plamo2 ), + MODEL_TENSOR.SSM_BETA_ALPHA: ( + "model.layers.{bid}.linear_attn.in_proj_ba", # qwen3next + ), + MODEL_TENSOR.TIME_MIX_W0: ( "model.layers.{bid}.attention.w0", # rwkv7 ), diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index f7a8c9841..67c7807e0 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -114,6 +114,7 @@ add_library(llama models/qwen3vl.cpp models/qwen3vl-moe.cpp models/qwen3moe.cpp + models/qwen3next.cpp models/refact.cpp models/rnd1.cpp models/rwkv6-base.cpp diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index f6e26245e..8571a2e02 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -32,6 +32,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_QWEN2VL, "qwen2vl" }, { LLM_ARCH_QWEN3, "qwen3" }, { LLM_ARCH_QWEN3MOE, "qwen3moe" }, + { LLM_ARCH_QWEN3NEXT, "qwen3next" }, { LLM_ARCH_QWEN3VL, "qwen3vl" }, { LLM_ARCH_QWEN3VLMOE, "qwen3vlmoe" }, { LLM_ARCH_PHI2, "phi2" }, @@ -829,6 +830,38 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, }, }, + { + LLM_ARCH_QWEN3NEXT, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" }, + { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" }, + { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" }, + { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, + { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" }, + { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" }, + { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" }, + { LLM_TENSOR_SSM_BETA_ALPHA, "blk.%d.ssm_ba" }, + { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" }, + { LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" }, + { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" }, + }, + }, { LLM_ARCH_QWEN3VL, { @@ -2556,6 +2589,7 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_SSM_X, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_SSM_DT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_SSM_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_SSM_BETA_ALPHA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_A1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, @@ -2754,6 +2788,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) { case LLM_ARCH_LFM2: case LLM_ARCH_LFM2MOE: case LLM_ARCH_NEMOTRON_H: + case LLM_ARCH_QWEN3NEXT: return true; default: return false; diff --git a/src/llama-arch.h b/src/llama-arch.h index 9ad3157bf..150646478 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -36,6 +36,7 @@ enum llm_arch { LLM_ARCH_QWEN2VL, LLM_ARCH_QWEN3, LLM_ARCH_QWEN3MOE, + LLM_ARCH_QWEN3NEXT, LLM_ARCH_QWEN3VL, LLM_ARCH_QWEN3VLMOE, LLM_ARCH_PHI2, @@ -381,6 +382,7 @@ enum llm_tensor { LLM_TENSOR_SSM_D, LLM_TENSOR_SSM_NORM, LLM_TENSOR_SSM_OUT, + LLM_TENSOR_SSM_BETA_ALPHA, // qwen3next LLM_TENSOR_TIME_MIX_W0, LLM_TENSOR_TIME_MIX_W1, LLM_TENSOR_TIME_MIX_W2, diff --git a/src/llama-context.cpp b/src/llama-context.cpp index 2aa6d52a2..a58914429 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -1,5 +1,6 @@ #include "llama-context.h" +#include "llama-arch.h" #include "llama-impl.h" #include "llama-batch.h" #include "llama-io.h" @@ -1386,6 +1387,9 @@ void llama_context::output_reorder() { // uint32_t llama_context::graph_max_nodes() const { + if (model.arch == LLM_ARCH_QWEN3NEXT) { + return std::max(8192u, 32u*model.n_tensors()); + } return std::max(1024u, 8u*model.n_tensors()); } diff --git a/src/llama-hparams.h b/src/llama-hparams.h index 9203af83b..c3a53be79 100644 --- a/src/llama-hparams.h +++ b/src/llama-hparams.h @@ -6,7 +6,7 @@ // bump if necessary #define LLAMA_MAX_LAYERS 512 -#define LLAMA_MAX_EXPERTS 384 // Kimi-K2 +#define LLAMA_MAX_EXPERTS 512 // Qwen3 Next enum llama_expert_gating_func_type { LLAMA_EXPERT_GATING_FUNC_TYPE_NONE = 0, diff --git a/src/llama-model.cpp b/src/llama-model.cpp index cba875f11..c2a545531 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -2,7 +2,6 @@ #include "llama-impl.h" #include "llama-mmap.h" -#include "llama-batch.h" #include "llama-cparams.h" #include "llama-model-loader.h" @@ -2225,6 +2224,29 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_QWEN3NEXT: + { + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false); + ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + // Load linear attention (gated delta net) parameters + ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv); + ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner); + ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state); + ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); + ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group); + + // Mark recurrent layers (linear attention layers) + for (uint32_t i = 0; i < hparams.n_layer; ++i) { + hparams.recurrent_layer_arr[i] = ((i + 1) % 4 != 0); // TODO: extract the magic 4 from "full_attention_interval" + } + + switch (hparams.n_layer) { + case 80: type = LLM_TYPE_80B_A3B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; default: throw std::runtime_error("unsupported model architecture"); } @@ -6415,6 +6437,74 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); } } break; + case LLM_ARCH_QWEN3NEXT: + { + 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); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED); + + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED); + } + + const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used; + + // Calculate dimensions from hyperparameters + const int64_t head_k_dim = hparams.ssm_d_state; + const int64_t head_v_dim = hparams.ssm_d_state; + const int64_t n_k_heads = hparams.ssm_n_group; + const int64_t n_v_heads = hparams.ssm_dt_rank; + const int64_t key_dim = head_k_dim * n_k_heads; + const int64_t value_dim = head_v_dim * n_v_heads; + const int64_t conv_dim = key_dim * 2 + value_dim; + + // Calculate projection sizes + const int64_t qkvz_dim = key_dim * 2 + value_dim * 2; + const int64_t ba_dim = n_v_heads * 2; + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0); + layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, 0); + + if (!hparams.is_recurrent(i)) { + // Attention layers + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head * 2 }, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0); + + // Q/K normalization for attention layers + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0); + } else { + // Linear attention (gated delta net) specific tensors + // Create tensors with calculated dimensions + layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), { n_embd, qkvz_dim }, 0); + layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), { hparams.ssm_d_conv, conv_dim }, 0); + layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), { hparams.ssm_dt_rank }, 0); + layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), { hparams.ssm_dt_rank }, 0); + layer.ssm_beta_alpha = create_tensor(tn(LLM_TENSOR_SSM_BETA_ALPHA, "weight", i), { n_embd, ba_dim }, 0); + layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), { head_v_dim }, 0); + layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), { value_dim, n_embd }, 0); + } + + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, 0); + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0); + + // Shared experts + layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), { n_embd }, 0); + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, hparams.n_ff_shexp }, 0); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, hparams.n_ff_shexp }, 0); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { hparams.n_ff_shexp, n_embd }, 0); + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -6685,6 +6775,7 @@ void llama_model::print_info() const { arch == LLM_ARCH_FALCON_H1 || arch == LLM_ARCH_PLAMO2 || arch == LLM_ARCH_GRANITE_HYBRID || + arch == LLM_ARCH_QWEN3NEXT || arch == LLM_ARCH_NEMOTRON_H) { LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv); LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner); @@ -7426,7 +7517,11 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const { case LLM_ARCH_PANGU_EMBED: { llm = std::make_unique(*this, params); - }break; + } break; + case LLM_ARCH_QWEN3NEXT: + { + llm = std::make_unique(*this, params); + } break; default: GGML_ABORT("fatal error"); } @@ -7653,6 +7748,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_COGVLM: case LLM_ARCH_PANGU_EMBED: case LLM_ARCH_AFMOE: + case LLM_ARCH_QWEN3NEXT: return LLAMA_ROPE_TYPE_NEOX; case LLM_ARCH_QWEN2VL: diff --git a/src/llama-model.h b/src/llama-model.h index f730c4954..f8342cf2c 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -113,6 +113,7 @@ enum llm_type { LLM_TYPE_16B_A1B, LLM_TYPE_21B_A3B, // Ernie MoE small LLM_TYPE_30B_A3B, + LLM_TYPE_80B_A3B, // Qwen3 Next LLM_TYPE_100B_A6B, LLM_TYPE_106B_A12B, // GLM-4.5-Air LLM_TYPE_230B_A10B, // Minimax M2 @@ -309,6 +310,9 @@ struct llama_layer { struct ggml_tensor * ssm_conv1d_b = nullptr; struct ggml_tensor * ssm_dt_b = nullptr; + // qwen3next + struct ggml_tensor * ssm_beta_alpha = nullptr; + // rwkv struct ggml_tensor * time_mix_w1 = nullptr; struct ggml_tensor * time_mix_w2 = nullptr; diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp index a56b2626a..0b23eaef3 100644 --- a/src/llama-quant.cpp +++ b/src/llama-quant.cpp @@ -681,7 +681,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: } LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str()); continue; - } else if (remapped_name != it.first) { + } + + if (remapped_name != it.first) { ggml_set_name(it.second.tensor, remapped_name.c_str()); LLAMA_LOG_DEBUG("%s: tensor %s remapped to %s\n", __func__, it.first.c_str(), ggml_get_name(it.second.tensor)); } @@ -726,13 +728,19 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: { const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin(); // attention layers have a non-zero number of kv heads - int32_t n_attn_layer = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0); + int32_t n_layer_attn = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0); if (llama_model_has_encoder(&model)) { - // now n_attn_layer is the number of attention layers in the encoder + // now n_layer_attn is the number of attention layers in the encoder // for each decoder block, there are 2 attention layers - n_attn_layer += 2 * model.hparams.dec_n_layer; + n_layer_attn += 2 * model.hparams.dec_n_layer; } - GGML_ASSERT((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected"); + + // note: for linear-attention models (such as Qwen3 Next) this is the number of linear layers + const int32_t n_layer_recr = std::count(model.hparams.recurrent_layer_arr.begin(), model.hparams.recurrent_layer_arr.end(), true); + + LLAMA_LOG_INFO("%s: n_layer_attn = %d, n_layer_recr = %d, pruned_attention_w = %d\n", __func__, n_layer_attn, n_layer_recr, pruned_attention_w); + + GGML_ASSERT((qs.n_attention_wv == n_layer_attn - pruned_attention_w - n_layer_recr) && "n_attention_wv is unexpected"); } size_t total_size_org = 0; diff --git a/src/models/models.h b/src/models/models.h index 5f019c59b..7ba225b47 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -2,8 +2,9 @@ #include "../llama-model.h" #include "../llama-graph.h" -#include "../llama-memory-recurrent.h" +// TODO: remove in follow-up PR - move to .cpp files +#include "../llama-memory-recurrent.h" #include struct llm_graph_context_mamba : public llm_graph_context { @@ -421,7 +422,56 @@ struct llm_build_qwen3vl : public llm_graph_context { struct llm_build_qwen3vlmoe : public llm_graph_context { llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params); }; +struct llm_build_qwen3next : public llm_graph_context_mamba { + llm_build_qwen3next(const llama_model & model, const llm_graph_params & params); +private: + ggml_tensor * build_layer_attn( + llm_graph_input_attn_kv * inp_attn, + ggml_tensor * cur, + ggml_tensor * inp_pos, + int il); + ggml_tensor * build_layer_attn_linear( + llm_graph_input_rs * inp, + ggml_tensor * cur, + ggml_tensor * causal_mask, + ggml_tensor * identity, + int il); + + ggml_tensor * build_layer_ffn( + ggml_tensor * cur, + int il); + + ggml_tensor * build_delta_net_recurrent( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + ggml_tensor * causal_mask, + ggml_tensor * identity, + int il); + + ggml_tensor * build_delta_net_chunking( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + ggml_tensor * causal_mask, + ggml_tensor * identity, + int il); + + ggml_tensor * build_norm_gated( + ggml_tensor * input, + ggml_tensor * weights, + ggml_tensor * gate, + int layer); + + const llama_model & model; +}; struct llm_build_qwen : public llm_graph_context { llm_build_qwen(const llama_model & model, const llm_graph_params & params); diff --git a/src/models/qwen3next.cpp b/src/models/qwen3next.cpp new file mode 100644 index 000000000..c8f1b5ec9 --- /dev/null +++ b/src/models/qwen3next.cpp @@ -0,0 +1,1042 @@ +#include "ggml.h" +#include "models.h" + +#define CHUNK_SIZE 64 + +llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_graph_params & params) : + llm_graph_context_mamba(params), model(model) { + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + cb(inpL, "model.embed_tokens", -1); + + auto * inp = build_inp_mem_hybrid(); + + ggml_tensor * inp_pos = build_inp_pos(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + ggml_tensor * causal_mask = + ggml_tri(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, ubatch.n_seq_tokens, ubatch.n_seq_tokens), 1.0f), + GGML_TRI_TYPE_LOWER); + + ggml_tensor * identity = ggml_diag(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, ubatch.n_seq_tokens), 1.0f)); + + ggml_build_forward_expand(gf, causal_mask); + ggml_build_forward_expand(gf, identity); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // Determine layer type and build appropriate attention mechanism + if (hparams.is_recurrent(il)) { + // Linear attention layer (gated delta net) + cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, il); + } else { + // Full attention layer + cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // Residual connection + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "attn_residual", il); + + // Save the tensor before post-attention norm for residual connection + ggml_tensor * ffn_residual = cur; + + // Post-attention norm + ggml_tensor * attn_post_norm = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il); + cb(attn_post_norm, "attn_post_norm", il); + + // FFN layer (MoE or dense) - without residual connection + cur = build_layer_ffn(attn_post_norm, il); + cb(cur, "ffn_out", il); + + // Residual connection for FFN - add to the tensor from before post_attention_layernorm + cur = ggml_add(ctx0, cur, ffn_residual); + cb(cur, "post_moe", il); + + // Input for next layer + inpL = cur; + } + cur = inpL; + + // Final norm + cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // LM head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + +ggml_tensor * llm_build_qwen3next::build_delta_net_chunking( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + ggml_tensor * causal_mask, + ggml_tensor * identity, + int il) { + GGML_ASSERT(ggml_is_contiguous(q)); + GGML_ASSERT(ggml_is_contiguous(k)); + GGML_ASSERT(ggml_is_contiguous(v)); + GGML_ASSERT(ggml_is_contiguous(g)); + GGML_ASSERT(ggml_is_contiguous(beta)); + GGML_ASSERT(ggml_is_contiguous(state)); + + const int64_t S_k = q->ne[0]; + const int64_t H_k = q->ne[1]; + const int64_t n_tokens = q->ne[2]; + const int64_t n_seqs = q->ne[3]; + + const int64_t S_v = v->ne[0]; + const int64_t H_v = v->ne[1]; + + GGML_ASSERT(v->ne[2] == n_tokens); + GGML_ASSERT(k->ne[2] == n_tokens); + GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); + GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs); + GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs); + + GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs); + + GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case + + // TODO: can this ever be false? + const bool use_qk_l2norm = true; + + if (use_qk_l2norm) { + const float eps_norm = hparams.f_norm_rms_eps; + + q = ggml_l2_norm(ctx0, q, eps_norm); + k = ggml_l2_norm(ctx0, k, eps_norm); + } + + const float scale = 1.0f / sqrtf(S_v); + + q = ggml_scale(ctx0, q, scale); + + beta = ggml_sigmoid(ctx0, beta); + + ggml_tensor * causal_diag_mask = ggml_add(ctx0, causal_mask, identity); + + cb(q, "q_in", il); + cb(k, "k_in", il); + cb(v, "v_in", il); + cb(beta, "beta_in", il); + cb(g, "g_in", il); + + q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs); + + beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3)); + state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); + + cb(q, "q_perm", il); + cb(k, "k_perm", il); + cb(v, "v_perm", il); + cb(beta, "beta_perm", il); + cb(g, "g_perm", il); + cb(state, "state_in", il); + + GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs); + GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs); + GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs); + + // Do padding + const int64_t chunk_size = CHUNK_SIZE; + + const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size; + const int64_t n_chunks = (n_tokens + pad) / chunk_size; + + q = ggml_pad(ctx0, q, 0, pad, 0, 0); + k = ggml_pad(ctx0, k, 0, pad, 0, 0); + v = ggml_pad(ctx0, v, 0, pad, 0, 0); + g = ggml_pad(ctx0, g, pad, 0, 0, 0); + beta = ggml_pad(ctx0, beta, 0, pad, 0, 0); + + cb(q, "q_pad", il); + cb(k, "k_pad", il); + cb(v, "v_pad", il); + cb(beta, "beta_pad", il); + cb(g, "g_pad", il); + + ggml_tensor * v_beta = ggml_mul(ctx0, v, beta); + ggml_tensor * k_beta = ggml_mul(ctx0, k, beta); + + cb(v_beta, "v_beta", il); + cb(k_beta, "k_beta", il); + + ggml_tensor * chunked_mask = + ggml_view_4d(ctx0, causal_mask, chunk_size, + chunk_size, causal_mask->ne[2], causal_mask->ne[3], + causal_mask->nb[1], causal_mask->nb[2], causal_mask->nb[3], 0); + + ggml_tensor * chunked_diag_mask = + ggml_view_4d(ctx0, causal_diag_mask, chunk_size, + chunk_size, causal_diag_mask->ne[2], causal_diag_mask->ne[3], + causal_diag_mask->nb[1], causal_diag_mask->nb[2], causal_diag_mask->nb[3], 0); + + ggml_tensor * chunked_identity = + ggml_view_4d(ctx0, identity, chunk_size, + chunk_size, identity->ne[2], identity->ne[3], + identity->nb[1], identity->nb[2], identity->nb[3], 0); + + q = ggml_cont_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs); + k = ggml_cont_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs); + k_beta = ggml_cont_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs); + v = ggml_cont_4d(ctx0, v, S_v, chunk_size, n_chunks, H_v * n_seqs); + v_beta = ggml_cont_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs); + + g = ggml_cont_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs); + beta = ggml_cont_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs); + + ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g); + + cb(g_cumsum, "g_cumsum", il); + + ggml_tensor * gcs_i = ggml_cont_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs); + ggml_tensor * gcs_j = ggml_cont_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs); + + ggml_tensor * gcs_j_broadcast = + ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs); + + ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i); + + cb(decay_mask, "decay_mask", il); + + decay_mask = ggml_mul(ctx0, decay_mask, chunked_diag_mask); + decay_mask = ggml_exp(ctx0, decay_mask); + decay_mask = ggml_mul(ctx0, decay_mask, chunked_diag_mask); + + ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta); + + ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask); + ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, chunked_mask)); + + cb(attn, "attn_pre_solve", il); + + ggml_tensor * attn_lower = ggml_mul(ctx0, attn, chunked_mask); + ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, chunked_identity, attn_lower), attn_lower); + + ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false); + attn = ggml_mul(ctx0, lin_solve, chunked_mask); + attn = ggml_add(ctx0, attn, chunked_identity); + + cb(attn, "attn_solved", il); + + v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn); + + ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum)); + ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t); + + ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp); + + cb(kbeta_gexp, "kbeta_gexp", il); + + ggml_tensor * k_cumdecay = + ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp))))); + + cb(k_cumdecay, "k_cumdecay", il); + + ggml_tensor * core_attn_out = nullptr; + ggml_tensor * new_state = ggml_dup(ctx0, state); + + cb(new_state, "new_state", il); + + for (int64_t chunk = 0; chunk < n_chunks; chunk++) { + auto chunkify = [=](ggml_tensor * t) { + return ggml_cont(ctx0, ggml_view_4d(ctx0, t, t->ne[0], chunk_size, 1, t->ne[3], + t->nb[1], t->nb[2], t->nb[3], t->nb[2] * chunk)); + }; + + auto chunkify_g = [=](ggml_tensor * t) { + return ggml_cont(ctx0, ggml_view_4d(ctx0, t, chunk_size, t->ne[1], 1, t->ne[3], + t->nb[1], t->nb[2], t->nb[3], t->nb[2] * chunk)); + }; + + ggml_tensor * k_chunk = chunkify(k); + ggml_tensor * q_chunk = chunkify(q); + ggml_tensor * v_chunk = chunkify(v); + + ggml_tensor * g_cs_chunk = chunkify_g(g_cumsum); + ggml_tensor * g_cs_chunk_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cs_chunk)); + + ggml_tensor * decay_mask_chunk = chunkify(decay_mask); + ggml_tensor * k_cumdecay_chunk = chunkify(k_cumdecay); + + ggml_tensor * gexp_chunk = ggml_exp(ctx0, g_cs_chunk_t); + + // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0) + attn = ggml_mul_mat(ctx0, k_chunk, q_chunk); + attn = ggml_mul(ctx0, attn, decay_mask_chunk); + attn = ggml_mul(ctx0, attn, ggml_add(ctx0, chunked_identity, chunked_mask)); + + ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs); + + // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state + ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk); + + // v_new = v_i - v_prime + ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime); + ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new)); + + // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state + ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk); + ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp); + + // core_attn_out[:, :, i] = attn_inter + attn @ v_new + ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn); + + ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn); + + core_attn_out = core_attn_out == nullptr ? core_attn_out_chunk : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 1); + + // g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1) + // g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp() + // key_gdiff = key * g_diff.unsqueeze(-1) + // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new + // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew + + ggml_tensor * g_cum_last = + ggml_cont(ctx0, ggml_view_4d(ctx0, g_cs_chunk_t, g_cs_chunk_t->ne[0], 1, g_cs_chunk_t->ne[2], g_cs_chunk_t->ne[3], + g_cs_chunk_t->nb[1], g_cs_chunk_t->nb[2], g_cs_chunk_t->nb[3], + g_cs_chunk_t->nb[0] * (g_cs_chunk_t->ne[1] - 1))); + + ggml_tensor * gexp_last = + ggml_reshape_4d(ctx0, ggml_exp(ctx0, g_cum_last), 1, 1, g_cum_last->ne[0] * g_cum_last->ne[2], g_cum_last->ne[3]); + + ggml_tensor * g_cum_last_3d = + ggml_reshape_3d(ctx0, g_cum_last, g_cum_last->ne[0], g_cum_last->ne[2], g_cum_last->ne[3]); + + ggml_tensor * g_cumsum_3d = ggml_reshape_3d(ctx0, g_cs_chunk, g_cs_chunk->ne[0], g_cs_chunk->ne[2], g_cs_chunk->ne[3]); + + ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum_3d, g_cum_last_3d)); + + ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff); + + ggml_tensor * key_gdiff = ggml_mul(ctx0, k_chunk, + ggml_reshape_4d(ctx0, g_diff_exp, 1, g_diff_exp->ne[0], g_diff_exp->ne[1], + g_diff_exp->ne[2] * g_diff_exp->ne[3])); + + ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff))); + + new_state = ggml_add(ctx0, + ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last, gexp_last->ne[0], gexp_last->ne[1], H_v, n_seqs)), + ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs)); + } + + core_attn_out = ggml_cont_4d(ctx0, core_attn_out, S_v, chunk_size * n_chunks, H_v, n_seqs); + + ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out, S_v, n_tokens, H_v, n_seqs, core_attn_out->nb[1], core_attn_out->nb[2], core_attn_out->nb[3], 0); + cb(output_tokens, "output_tokens", il); + + // flatten output + ggml_tensor * flat_output = + ggml_cont_1d(ctx0, ggml_permute(ctx0, output_tokens, 0, 2, 1, 3), S_v * H_v * n_tokens * n_seqs); + + ggml_tensor * flat_state = ggml_cont_1d(ctx0, new_state, S_v * S_v * H_v * n_seqs); + + return ggml_concat(ctx0, flat_output, flat_state, 0); +} + +ggml_tensor * llm_build_qwen3next::build_delta_net_recurrent( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + ggml_tensor * causal_mask, + ggml_tensor * identity, + int il) { + GGML_ASSERT(ggml_is_contiguous(q)); + GGML_ASSERT(ggml_is_contiguous(k)); + GGML_ASSERT(ggml_is_contiguous(v)); + GGML_ASSERT(ggml_is_contiguous(g)); + GGML_ASSERT(ggml_is_contiguous(beta)); + GGML_ASSERT(ggml_is_contiguous(state)); + + const int64_t S_k = q->ne[0]; + const int64_t H_k = q->ne[1]; + const int64_t n_tokens = q->ne[2]; + const int64_t n_seqs = q->ne[3]; + + const int64_t S_v = v->ne[0]; + const int64_t H_v = v->ne[1]; + + GGML_ASSERT(v->ne[2] == n_tokens); + GGML_ASSERT(k->ne[2] == n_tokens); + GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); + GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs); + GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs); + + GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs); + + GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case + + // TODO: can this ever be false? + const bool use_qk_l2norm = true; + + if (use_qk_l2norm) { + const float eps_norm = hparams.f_norm_rms_eps; + + q = ggml_l2_norm(ctx0, q, eps_norm); + k = ggml_l2_norm(ctx0, k, eps_norm); + } + + const float scale = 1.0f / sqrtf(S_v); + + q = ggml_scale(ctx0, q, scale); + + beta = ggml_sigmoid(ctx0, beta); + + ggml_tensor * causal_diag_mask = ggml_add(ctx0, causal_mask, identity); + + cb(q, "q_in", il); + cb(k, "k_in", il); + cb(v, "v_in", il); + cb(beta, "beta_in", il); + cb(g, "g_in", il); + + q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs); + + beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3)); + state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); + + cb(q, "q_perm", il); + cb(k, "k_perm", il); + cb(v, "v_perm", il); + cb(beta, "beta_perm", il); + cb(g, "g_perm", il); + cb(state, "state_in", il); + + GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs); + GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs); + GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs); + + ggml_tensor * v_beta = ggml_mul(ctx0, v, beta); + ggml_tensor * k_beta = ggml_mul(ctx0, k, beta); + + ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g); + + cb(k_beta, "k_beta", il); + cb(v_beta, "v_beta", il); + cb(g_cumsum, "g_cumsum", il); + + ggml_tensor * gcs_i = ggml_cont_4d(ctx0, g_cumsum, n_tokens, 1, H_v, n_seqs); // [chunk_size, 1, n_tokens, n_seqs] + ggml_tensor * gcs_j = ggml_cont_4d(ctx0, g_cumsum, 1, n_tokens, H_v, n_seqs); // [1, chunk_size, n_tokens, n_seqs] + + // Broadcast both tensors to [chunk_size, chunk_size, H_v, n_seqs] + // ggml_tensor * gcs_i_broadcast = + // ggml_repeat_4d(ctx0, gcs_i, GGML_DELTA_NET_CHUNK, GGML_DELTA_NET_CHUNK, num_chunks * H_v, + // n_seqs); // [chunk_size, 1, H_v, n_seqs] -> [chunk_size, chunk_size, H_v, n_seqs] + // Don't need this, this one will get auto-broadcast + ggml_tensor * gcs_j_broadcast = + ggml_repeat_4d(ctx0, gcs_j, n_tokens, n_tokens, H_v, n_seqs); // [1, chunk_size, H_v, n_seqs] -> [chunk_size, chunk_size, H_v, n_seqs] + + ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i); + + // Apply lower triangular mask to ensure attention is causal (only past tokens influence current) + decay_mask = ggml_mul(ctx0, decay_mask, causal_diag_mask); + // Apply exponential to get the decay mask values + decay_mask = ggml_exp(ctx0, decay_mask); + // Apply lower triangular mask again to ensure only lower triangular values remain + decay_mask = ggml_mul(ctx0, decay_mask, causal_diag_mask); + + cb(decay_mask, "decay_mask", il); + + // attn = -((k_beta @ key.transpose(-1, -2)) * decay_mask).masked_fill(mask, 0) + ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta); + + cb(kmulkbeta, "kmulkbeta", il); + + ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask); + ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask)); + + cb(attn, "attn_pre_rec", il); + + // for i in range(1, chunk_size): + // row = attn[..., i, :i].clone() + // sub = attn[..., :i, :i].clone() + // attn[..., i, :i] = row + (row.unsqueeze(-1) * sub).sum(-2) + // attn = attn + torch.eye(chunk_size, dtype=attn.dtype, device=attn.device) + // + // We reduce this to a linear triangular solve: AX = B, where B = attn, A = I - tril(A) + ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask); + ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower); + + ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false); + attn = ggml_mul(ctx0, lin_solve, causal_mask); + attn = ggml_add(ctx0, attn, identity); + + // value = attn @ v_beta + v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn); + + cb(v, "value_beta", il); + + // k_cumdecay = attn @ (k_beta * g.exp().unsqueeze(-1)) + ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum)); + ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t); + + cb(gexp, "g_cum_exp", il); + + ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp); + + cb(kbeta_gexp, "kbeta_gexp", il); + + ggml_tensor * k_cumdecay = + ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp))))); + + cb(k_cumdecay, "k_cumdecay", il); + + // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0) + attn = ggml_mul_mat(ctx0, k, q); + attn = ggml_mul(ctx0, attn, decay_mask); + attn = ggml_mul(ctx0, attn, ggml_add(ctx0, identity, causal_mask)); + + cb(attn, "attn_decay_key", il); + + ggml_tensor * state_t = ggml_cont(ctx0, ggml_transpose(ctx0, state)); + + // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state + ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay); + + cb(v_prime, "v_prime", il); + + // v_new = v_i - v_prime + ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v, v_prime), v_prime); + + ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new)); + + cb(v_new, "v_new", il); + + // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state + ggml_tensor * q_g_exp = ggml_mul(ctx0, q, gexp); + ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp); + + cb(attn_inter, "attn_inter", il); + + // core_attn_out[:, :, i] = attn_inter + attn @ v_new + ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn); + + cb(v_attn, "v_attn", il); + + ggml_tensor * core_attn_out = ggml_add(ctx0, attn_inter, v_attn); + + cb(core_attn_out, "core_attn_out", il); + + // g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1) + // g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp() + // key_gdiff = key * g_diff.unsqueeze(-1) + // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new + // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew + + ggml_tensor * g_cum_last = + ggml_cont(ctx0, ggml_view_4d(ctx0, g_cumsum_t, g_cumsum_t->ne[0], 1, g_cumsum_t->ne[2], g_cumsum_t->ne[3], + g_cumsum_t->nb[1], g_cumsum_t->nb[2], g_cumsum_t->nb[3], + g_cumsum_t->nb[0] * (g_cumsum_t->ne[1] - 1))); + + cb(g_cum_last, "g_cum_last", il); + + ggml_tensor * gexp_last = + ggml_reshape_4d(ctx0, ggml_exp(ctx0, g_cum_last), 1, 1, g_cum_last->ne[0] * g_cum_last->ne[2], g_cum_last->ne[3]); + + cb(gexp_last, "gexp_last", il); + + ggml_tensor * g_cum_last_3d = + ggml_reshape_3d(ctx0, g_cum_last, g_cum_last->ne[0], g_cum_last->ne[2], g_cum_last->ne[3]); + + cb(g_cum_last_3d, "g_cum_last_3d", il); + + ggml_tensor * g_cumsum_3d = ggml_reshape_3d(ctx0, g_cumsum, g_cumsum->ne[0], g_cumsum->ne[2], g_cumsum->ne[3]); + + cb(g_cumsum_3d, "g_cumsum_3d", il); + + ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum_3d, g_cum_last_3d)); + + cb(g_diff, "g_diff", il); + + ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff); + + cb(g_diff_exp, "g_diff_exp", il); + + ggml_tensor * key_gdiff = ggml_mul(ctx0, k, + ggml_reshape_4d(ctx0, g_diff_exp, 1, g_diff_exp->ne[0], g_diff_exp->ne[1], + g_diff_exp->ne[2] * g_diff_exp->ne[3])); + + cb(key_gdiff, "key_gdiff", il); + + ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff))); + + cb(kgdmulvnew, "kgdmulvnew", il); + + state = ggml_add(ctx0, ggml_mul(ctx0, state, gexp_last), kgdmulvnew); + + cb(state, "new_state", il); + + // flatten output + ggml_tensor * flat_output = + ggml_cont_1d(ctx0, ggml_permute(ctx0, core_attn_out, 0, 2, 1, 3), S_v * H_v * n_tokens * n_seqs); + + ggml_tensor * flat_state = ggml_cont_1d(ctx0, state, S_v * S_v * H_v * n_seqs); + + return ggml_concat(ctx0, flat_output, flat_state, 0); +} + +ggml_tensor * llm_build_qwen3next::build_norm_gated( + ggml_tensor * input, + ggml_tensor * weights, + ggml_tensor * gate, + int layer) { + ggml_tensor * normalized = build_norm(input, weights, nullptr, LLM_NORM_RMS, layer); + ggml_tensor * gated_silu = ggml_silu(ctx0, gate); + + return ggml_mul(ctx0, normalized, gated_silu); +} + +ggml_tensor * llm_build_qwen3next::build_layer_attn( + llm_graph_input_attn_kv * inp, + ggml_tensor * cur, + ggml_tensor * inp_pos, + int il) { + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention + + // Qwen3Next uses a single Q projection that outputs query + gate + ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur_full, "Qcur_full", il); + + Qcur_full = ggml_reshape_4d(ctx0, Qcur_full, n_embd_head * 2, n_head, n_tokens, 1); + + // Split Q projection into query and gate + // The split should be along dimension 0 (the feature dimension) + ggml_tensor * Qcur = ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1, + Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0); + ggml_tensor * gate = + ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1, + Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], n_embd_head * ggml_element_size(Qcur_full)); + cb(Qcur, "Qcur", il); + cb(gate, "gate", il); + + // Now reshape Qcur to [n_embd_head, n_head, n_tokens] for multi-head attention + Qcur = ggml_cont_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + cb(Qcur, "Qcur_reshaped", il); + + // Apply Q normalization + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + // Apply K normalization + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + // Reshape gate to [n_embd, n_tokens] for the sigmoid gating (flatten the heads) + gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens); + cb(gate, "gate_reshaped", il); + + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + // Apply RoPE + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, + freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + // Attention computation + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + cur = build_attn(inp, + nullptr, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_pregate", il); + + ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate); + cb(gate_sigmoid, "gate_sigmoid", il); + + cur = ggml_mul(ctx0, cur, gate_sigmoid); + cb(cur, "attn_gated", il); + + cur = build_lora_mm(model.layers[il].wo, cur); + cb(cur, "attn_output", il); + + return cur; +} + +ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( + llm_graph_input_rs * inp, + ggml_tensor * cur, + ggml_tensor * causal_mask, + ggml_tensor * identity, + int il) { + const auto * mctx_cur = inp->mctx; + + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t n_seqs = ubatch.n_seqs; + const int64_t head_k_dim = hparams.ssm_d_state; + const int64_t num_k_heads = hparams.ssm_n_group; + const int64_t num_v_heads = hparams.ssm_dt_rank; + const int64_t head_v_dim = d_inner / num_v_heads; + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + const auto kv_head = mctx_cur->get_head(); + + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(ubatch.equal_seqs()); + GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); + + // Input projections + ggml_tensor * mixed_qkvz = build_lora_mm(model.layers[il].ssm_in, cur); + cb(mixed_qkvz, "linear_attn_mixed_qkvz", il); + + ggml_tensor * mixed_ba = build_lora_mm(model.layers[il].ssm_beta_alpha, cur); + cb(mixed_ba, "linear_attn_mixed_ba", il); + + int64_t qkvz_new_dim = 2 * head_k_dim + 2 * head_v_dim * (num_v_heads / num_k_heads); + ggml_tensor * mixed_qkvz_reshaped = ggml_cont_4d(ctx0, mixed_qkvz, qkvz_new_dim, num_k_heads, n_seq_tokens, n_seqs); + + // Reshape mixed_ba: [batch, seq_len, hidden_size] -> [batch, seq_len, num_k_heads, 2*num_v_heads/num_k_heads] + int64_t ba_new_dim = 2 * num_v_heads / num_k_heads; + ggml_tensor * mixed_ba_reshaped = ggml_cont_4d(ctx0, mixed_ba, ba_new_dim, num_k_heads, n_seq_tokens, n_seqs); + + // Split mixed_ba into b and a (beta and alpha parameters) + int64_t split_sizes_ba[2] = { + num_v_heads / num_k_heads, // beta size + num_v_heads / num_k_heads // alpha size + }; + + ggml_tensor * b = ggml_view_4d(ctx0, mixed_ba_reshaped, split_sizes_ba[0], num_k_heads, n_seq_tokens, n_seqs, + mixed_ba_reshaped->nb[1], mixed_ba_reshaped->nb[2], mixed_ba_reshaped->nb[3], 0); + cb(b, "b", il); + + ggml_tensor * a = ggml_view_4d(ctx0, mixed_ba_reshaped, split_sizes_ba[1], num_k_heads, n_seq_tokens, n_seqs, + mixed_ba_reshaped->nb[1], mixed_ba_reshaped->nb[2], mixed_ba_reshaped->nb[3], + split_sizes_ba[0] * ggml_element_size(mixed_ba_reshaped)); + cb(a, "a", il); + + // Reshape b and a to merge head dimensions: [batch, seq_len, num_k_heads, num_v_heads/num_k_heads] -> [batch, seq_len, num_v_heads] + ggml_tensor * beta = ggml_cont_3d(ctx0, b, num_v_heads, n_seq_tokens, n_seqs); + ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs); + + GGML_ASSERT(ggml_nelements(beta) + ggml_nelements(alpha) == ggml_nelements(mixed_ba)); + + ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt); + ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased); + cb(alpha_softplus, "a_softplus", il); + ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus + cb(gate, "gate", il); + + // Split mixed_qkvz into query, key, value, z + int64_t split_sizes_qkvz[4] = { + head_k_dim, // query size + head_k_dim, // key size + head_v_dim * num_v_heads / num_k_heads, // value size + head_v_dim * num_v_heads / num_k_heads // z size + }; + + ggml_tensor * query = + ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[0], num_k_heads, n_seq_tokens, n_seqs, + mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3], 0); + cb(query, "q", il); + + ggml_tensor * key = ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[1], num_k_heads, n_seq_tokens, n_seqs, + mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3], + split_sizes_qkvz[0] * sizeof(float)); + cb(key, "k", il); + + ggml_tensor * value = + ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[2], num_k_heads, n_seq_tokens, n_seqs, + mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3], + (split_sizes_qkvz[0] + split_sizes_qkvz[1]) * sizeof(float)); + cb(value, "v", il); + + ggml_tensor * z = ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[3], num_k_heads, n_seq_tokens, n_seqs, + mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3], + (split_sizes_qkvz[0] + split_sizes_qkvz[1] + split_sizes_qkvz[2]) * sizeof(float)); + cb(z, "z", il); + + GGML_ASSERT(ggml_nelements(query) + ggml_nelements(key) + ggml_nelements(value) + ggml_nelements(z) == + ggml_nelements(mixed_qkvz)); + + // After creating query, key, and value_reshaped, reshape each to flatten the head dimensions + // query: [head_k_dim, num_k_heads, n_tokens, n_seqs] -> [head_k_dim * num_k_heads, n_tokens, n_seqs] + ggml_tensor * query_flat = ggml_cont_3d(ctx0, query, head_k_dim * num_k_heads, n_seq_tokens, n_seqs); + cb(query_flat, "query_flat", il); + + // key: [head_k_dim, num_k_heads, n_tokens, n_seqs] -> [head_k_dim * num_k_heads, n_tokens, n_seqs] + ggml_tensor * key_flat = ggml_cont_3d(ctx0, key, head_k_dim * num_k_heads, n_seq_tokens, n_seqs); + cb(key_flat, "key_flat", il); + + // value_reshaped: [head_v_dim, num_v_heads, n_tokens, n_seqs] -> [head_v_dim * num_v_heads, n_tokens, n_seqs] + ggml_tensor * value_flat = ggml_cont_3d(ctx0, value, head_v_dim * num_v_heads, n_seq_tokens, n_seqs); + cb(value_flat, "value_flat", il); + + // Get convolution states from cache + ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); + ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); + + // bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state(); + + // Build the convolution states tensor + ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs); + cb(conv_states, "conv_states", il); + + // Now concatenate along the feature dimension (dim 0) to get [conv_dim, n_tokens, n_seqs] + ggml_tensor * qkv_mixed = ggml_concat(ctx0, query_flat, key_flat, 0); + qkv_mixed = ggml_concat(ctx0, qkv_mixed, value_flat, 0); + cb(qkv_mixed, "qkv_mixed", il); + + qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3); + cb(qkv_mixed, "qkv_mixed_permuted", il); + + // Calculate the total conv dimension + int64_t qkv_dim = head_k_dim * num_k_heads * 2 + head_v_dim * num_v_heads; + + // Calculate convolution kernel size + ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d; + const int64_t conv_kernel_size = conv_kernel->ne[0]; + const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state; + conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs); + cb(conv_states, "conv_states_reshaped", il); + + ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0); + cb(conv_input, "conv_input", il); + + // Update convolution state cache + // Extract the last (conv_kernel_size - 1) states from conv_input + ggml_tensor * last_conv_states = + ggml_view_3d(ctx0, conv_input, conv_kernel_size - 1, conv_channels, n_seqs, conv_input->nb[1], + conv_input->nb[2], (conv_input->ne[0] - conv_states->ne[0]) * ggml_element_size(conv_input)); + cb(last_conv_states, "last_conv_states", il); + + ggml_tensor * state_update_target = + ggml_view_1d(ctx0, conv_states_all, (conv_kernel_size - 1) * conv_channels * n_seqs, + kv_head * (conv_kernel_size - 1) * conv_channels * ggml_element_size(conv_states_all)); + cb(state_update_target, "state_update_target", il); + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target)); + cb(conv_states_all, "conv_states_updated", il); + + // Apply SSM convolution + ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel); + cb(conv_output_proper, "conv_output_raw", il); + + conv_output_proper = ggml_cont(ctx0, ggml_transpose(ctx0, conv_output_proper)); + cb(conv_output_proper, "conv_output_pre_silu", il); + + ggml_tensor * conv_output_silu = ggml_silu(ctx0, conv_output_proper); + cb(conv_output_silu, "conv_output_silu", il); + + ggml_tensor * conv_qkv_mix = + ggml_cont_2d(ctx0, ggml_transpose(ctx0, conv_output_silu), qkv_dim, n_seq_tokens * n_seqs); + cb(conv_qkv_mix, "conv_qkv_mix", il); + + // Extract the convolved Q, K, V from conv_output + ggml_tensor * q_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, conv_qkv_mix->nb[1], 0); + cb(q_conv, "q_conv", il); + ggml_tensor * k_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, conv_qkv_mix->nb[1], + head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); + cb(k_conv, "k_conv", il); + ggml_tensor * v_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, conv_qkv_mix->nb[1], + 2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); + cb(v_conv, "v_conv", il); + + // Unsqueeze them + q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); + k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); + v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs); + + beta = ggml_cont_4d(ctx0, b, num_v_heads, 1, n_seq_tokens, n_seqs); + + ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs); + state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs); + cb(state, "state_predelta", il); + + // if head keys and value keys are different, repeat to force tensors into matching shapes + if (num_k_heads != num_v_heads) { + GGML_ASSERT(num_v_heads % num_k_heads == 0); + int64_t repeat_factor = num_v_heads / num_k_heads; + + // repeat interleave: reshape to (repeat part, 1, remaining part), do repeat, then reshape back + ggml_tensor * q_reshaped = ggml_reshape_3d(ctx0, q_conv, head_k_dim, 1, num_k_heads * n_seq_tokens * n_seqs); + ggml_tensor * k_reshaped = ggml_reshape_3d(ctx0, k_conv, head_k_dim, 1, num_k_heads * n_seq_tokens * n_seqs); + + // Repeat along the third dimension (the new dimension with size 1) + ggml_tensor * q_repeated = + ggml_repeat_4d(ctx0, q_reshaped, head_k_dim, repeat_factor, num_k_heads * n_seq_tokens * n_seqs, 1); + ggml_tensor * k_repeated = + ggml_repeat_4d(ctx0, k_reshaped, head_k_dim, repeat_factor, num_k_heads * n_seq_tokens * n_seqs, 1); + + // Reshape back to merge the head and repeat dimensions + // From [head_dim, num_k_heads, repeat_factor, n_seq_tokens * n_seqs] + // Back to [head_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs] + q_conv = ggml_reshape_4d(ctx0, q_repeated, head_k_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs); + k_conv = ggml_reshape_4d(ctx0, k_repeated, head_k_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs); + } + + cb(q_conv, "q_conv_predelta", il); + cb(k_conv, "k_conv_predelta", il); + cb(v_conv, "v_conv_predelta", il); + + // Choose between build_delta_net_chunking and build_delta_net_recurrent based on n_tokens + ggml_tensor * attn_out = n_seq_tokens > CHUNK_SIZE ? + build_delta_net_chunking (q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, il) : + build_delta_net_recurrent(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, il); + cb(attn_out, "attn_out", il); + + // The tensors were concatenated 1d, so we need to extract them 1d as well + const int64_t output_flat_size = head_v_dim * num_v_heads * n_seq_tokens * n_seqs; + ggml_tensor * attn_out_1d = ggml_view_1d(ctx0, attn_out, output_flat_size, 0); + cb(attn_out_1d, "attn_out_1d", il); + + ggml_tensor * attn_out_final = ggml_cont_4d(ctx0, attn_out_1d, head_v_dim, num_v_heads, n_seq_tokens, n_seqs); + cb(attn_out_final, "attn_out_reshaped", il); + + // Extract the state part (second part of the concatenated tensor) + // State starts after n_tokens elements along dimension 1 + const int64_t state_flat_size = head_v_dim * head_v_dim * num_v_heads * n_seqs; + + ggml_tensor * state_1d = + ggml_view_1d(ctx0, attn_out, state_flat_size, output_flat_size * ggml_element_size(attn_out)); + cb(state_1d, "state_1d", il); + + // Update the recurrent states + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, state_1d, + ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs, + kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all)))); + + GGML_ASSERT(ggml_nelements(attn_out_1d) + ggml_nelements(state_1d) == ggml_nelements(attn_out)); + + // Reshape both attn_out_final and z to 2D tensors for normalization + // attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] + ggml_tensor * attn_out_2d_final = + ggml_cont_2d(ctx0, attn_out_final, head_v_dim, num_v_heads * n_seq_tokens * n_seqs); + + // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] + ggml_tensor * z_2d = ggml_cont_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs); + + // Apply gated normalization: self.norm(core_attn_out, z) + ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il); + + // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim] + ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs); + cb(final_output, "final_output", il); + + // Output projection + cur = build_lora_mm(model.layers[il].ssm_out, final_output); + cb(cur, "linear_attn_out", il); + + // Reshape back to original dimensions + cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs); + return cur; +} + +ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int il) { + // Check if this is an MoE layer + if (model.layers[il].ffn_gate_inp != nullptr) { + // 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, + nullptr, + n_expert, n_expert_used, LLM_FFN_SILU, + true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il); + cb(moe_out, "ffn_moe_out", il); + + // Add shared experts if present - following Qwen3Next reference implementation + if (model.layers[il].ffn_up_shexp != nullptr) { + ggml_tensor * ffn_shexp = + build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + // Apply shared expert gating as in the reference implementation + // The shared expert has its own gate that is sigmoided + // Note: ffn_gate_inp_shexp is the shared expert gate (outputs 1 value per token) + ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur); + cb(shared_gate, "shared_expert_gate", il); + + // Apply sigmoid to the gate + shared_gate = ggml_sigmoid(ctx0, shared_gate); + cb(shared_gate, "shared_expert_gate_sigmoid", il); + + // The gate needs to be broadcast to match the dimensions of ffn_shexp + // ffn_shexp is [n_embd, n_tokens, 1, 1] and shared_gate is [1, n_tokens, 1, 1] + // We need to repeat the gate along the feature dimension + shared_gate = ggml_repeat(ctx0, shared_gate, ffn_shexp); + cb(shared_gate, "shared_expert_gate_broadcast", il); + + // Apply the gate to the shared expert output + ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate); + cb(ffn_shexp, "ffn_shexp_gated", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } else { + cur = moe_out; + } + } else { + // Dense FFN branch (not currently used I believe) + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + return cur; +} From ddf9f94389a614ce005347f1c3f60ce477df1be1 Mon Sep 17 00:00:00 2001 From: Fredrik Hultin Date: Fri, 28 Nov 2025 12:57:04 +0100 Subject: [PATCH 2/2] server : add Anthropic Messages API support (#17570) * server : add Anthropic Messages API support * remove -@pytest.mark.slow from tool calling/jinja tests * server : remove unused code and slow/skip on test_anthropic_vision_base64_with_multimodal_model in test_anthropic_api.py * server : removed redundant n field logic in anthropic_params_from_json * server : use single error object instead of error_array in streaming response handler for /v1/chat/completions and use unordered_set instead of set in to_json_anthropic_stream() * server : refactor Anthropic API to use OAI conversion * make sure basic test always go first * clean up * clean up api key check, add test --------- Co-authored-by: Xuan Son Nguyen --- tools/server/README.md | 72 ++ tools/server/server-common.cpp | 242 +++++- tools/server/server-common.h | 8 +- tools/server/server-http.cpp | 21 +- tools/server/server-task.cpp | 304 ++++++- tools/server/server-task.h | 47 +- tools/server/server.cpp | 97 ++- tools/server/tests/conftest.py | 6 + tools/server/tests/unit/test_basic.py | 6 - .../tests/unit/test_compat_anthropic.py | 807 ++++++++++++++++++ tools/server/tests/unit/test_security.py | 13 + 11 files changed, 1553 insertions(+), 70 deletions(-) create mode 100644 tools/server/tests/unit/test_compat_anthropic.py diff --git a/tools/server/README.md b/tools/server/README.md index b7fc565ec..f42bc7921 100644 --- a/tools/server/README.md +++ b/tools/server/README.md @@ -7,6 +7,7 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp. **Features:** * LLM inference of F16 and quantized models on GPU and CPU * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes + * [Anthropic Messages API](https://docs.anthropic.com/en/api/messages) compatible chat completions * Reranking endpoint (https://github.com/ggml-org/llama.cpp/pull/9510) * Parallel decoding with multi-user support * Continuous batching @@ -1352,6 +1353,77 @@ See [OpenAI Embeddings API documentation](https://platform.openai.com/docs/api-r }' ``` +### POST `/v1/messages`: Anthropic-compatible Messages API + +Given a list of `messages`, returns the assistant's response. Streaming is supported via Server-Sent Events. While no strong claims of compatibility with the Anthropic API spec are made, in our experience it suffices to support many apps. + +*Options:* + +See [Anthropic Messages API documentation](https://docs.anthropic.com/en/api/messages). Tool use requires `--jinja` flag. + +`model`: Model identifier (required) + +`messages`: Array of message objects with `role` and `content` (required) + +`max_tokens`: Maximum tokens to generate (default: 4096) + +`system`: System prompt as string or array of content blocks + +`temperature`: Sampling temperature 0-1 (default: 1.0) + +`top_p`: Nucleus sampling (default: 1.0) + +`top_k`: Top-k sampling + +`stop_sequences`: Array of stop sequences + +`stream`: Enable streaming (default: false) + +`tools`: Array of tool definitions (requires `--jinja`) + +`tool_choice`: Tool selection mode (`{"type": "auto"}`, `{"type": "any"}`, or `{"type": "tool", "name": "..."}`) + +*Examples:* + +```shell +curl http://localhost:8080/v1/messages \ + -H "Content-Type: application/json" \ + -H "x-api-key: your-api-key" \ + -d '{ + "model": "gpt-4", + "max_tokens": 1024, + "system": "You are a helpful assistant.", + "messages": [ + {"role": "user", "content": "Hello!"} + ] + }' +``` + +### POST `/v1/messages/count_tokens`: Token Counting + +Counts the number of tokens in a request without generating a response. + +Accepts the same parameters as `/v1/messages`. The `max_tokens` parameter is not required. + +*Example:* + +```shell +curl http://localhost:8080/v1/messages/count_tokens \ + -H "Content-Type: application/json" \ + -d '{ + "model": "gpt-4", + "messages": [ + {"role": "user", "content": "Hello!"} + ] + }' +``` + +*Response:* + +```json +{"input_tokens": 10} +``` + ## More examples ### Interactive mode diff --git a/tools/server/server-common.cpp b/tools/server/server-common.cpp index 18328f3af..0bbc4e858 100644 --- a/tools/server/server-common.cpp +++ b/tools/server/server-common.cpp @@ -725,7 +725,6 @@ std::vector tokenize_input_prompts(const llama_vocab * vocab, mtm return result; } - // // OAI utils // @@ -1048,6 +1047,222 @@ json oaicompat_chat_params_parse( return llama_params; } +json convert_anthropic_to_oai(const json & body) { + json oai_body; + + // Convert system prompt + json oai_messages = json::array(); + auto system_param = json_value(body, "system", json()); + if (!system_param.is_null()) { + std::string system_content; + + if (system_param.is_string()) { + system_content = system_param.get(); + } else if (system_param.is_array()) { + for (const auto & block : system_param) { + if (json_value(block, "type", std::string()) == "text") { + system_content += json_value(block, "text", std::string()); + } + } + } + + oai_messages.push_back({ + {"role", "system"}, + {"content", system_content} + }); + } + + // Convert messages + if (!body.contains("messages")) { + throw std::runtime_error("'messages' is required"); + } + const json & messages = body.at("messages"); + if (messages.is_array()) { + for (const auto & msg : messages) { + std::string role = json_value(msg, "role", std::string()); + + if (!msg.contains("content")) { + if (role == "assistant") { + continue; + } + oai_messages.push_back(msg); + continue; + } + + const json & content = msg.at("content"); + + if (content.is_string()) { + oai_messages.push_back(msg); + continue; + } + + if (!content.is_array()) { + oai_messages.push_back(msg); + continue; + } + + json tool_calls = json::array(); + json converted_content = json::array(); + json tool_results = json::array(); + bool has_tool_calls = false; + + for (const auto & block : content) { + std::string type = json_value(block, "type", std::string()); + + if (type == "text") { + converted_content.push_back(block); + } else if (type == "image") { + json source = json_value(block, "source", json::object()); + std::string source_type = json_value(source, "type", std::string()); + + if (source_type == "base64") { + std::string media_type = json_value(source, "media_type", std::string("image/jpeg")); + std::string data = json_value(source, "data", std::string()); + std::ostringstream ss; + ss << "data:" << media_type << ";base64," << data; + + converted_content.push_back({ + {"type", "image_url"}, + {"image_url", { + {"url", ss.str()} + }} + }); + } else if (source_type == "url") { + std::string url = json_value(source, "url", std::string()); + converted_content.push_back({ + {"type", "image_url"}, + {"image_url", { + {"url", url} + }} + }); + } + } else if (type == "tool_use") { + tool_calls.push_back({ + {"id", json_value(block, "id", std::string())}, + {"type", "function"}, + {"function", { + {"name", json_value(block, "name", std::string())}, + {"arguments", json_value(block, "input", json::object()).dump()} + }} + }); + has_tool_calls = true; + } else if (type == "tool_result") { + std::string tool_use_id = json_value(block, "tool_use_id", std::string()); + + auto result_content = json_value(block, "content", json()); + std::string result_text; + if (result_content.is_string()) { + result_text = result_content.get(); + } else if (result_content.is_array()) { + for (const auto & c : result_content) { + if (json_value(c, "type", std::string()) == "text") { + result_text += json_value(c, "text", std::string()); + } + } + } + + tool_results.push_back({ + {"role", "tool"}, + {"tool_call_id", tool_use_id}, + {"content", result_text} + }); + } + } + + if (!converted_content.empty() || has_tool_calls) { + json new_msg = {{"role", role}}; + if (!converted_content.empty()) { + new_msg["content"] = converted_content; + } else if (has_tool_calls) { + new_msg["content"] = ""; + } + if (!tool_calls.empty()) { + new_msg["tool_calls"] = tool_calls; + } + oai_messages.push_back(new_msg); + } + + for (const auto & tool_msg : tool_results) { + oai_messages.push_back(tool_msg); + } + } + } + + oai_body["messages"] = oai_messages; + + // Convert tools + if (body.contains("tools")) { + const json & tools = body.at("tools"); + if (tools.is_array()) { + json oai_tools = json::array(); + for (const auto & tool : tools) { + oai_tools.push_back({ + {"type", "function"}, + {"function", { + {"name", json_value(tool, "name", std::string())}, + {"description", json_value(tool, "description", std::string())}, + {"parameters", tool.contains("input_schema") ? tool.at("input_schema") : json::object()} + }} + }); + } + oai_body["tools"] = oai_tools; + } + } + + // Convert tool_choice + if (body.contains("tool_choice")) { + const json & tc = body.at("tool_choice"); + if (tc.is_object()) { + std::string type = json_value(tc, "type", std::string()); + if (type == "auto") { + oai_body["tool_choice"] = "auto"; + } else if (type == "any" || type == "tool") { + oai_body["tool_choice"] = "required"; + } + } + } + + // Convert stop_sequences to stop + if (body.contains("stop_sequences")) { + oai_body["stop"] = body.at("stop_sequences"); + } + + // Handle max_tokens (required in Anthropic, but we're permissive) + if (body.contains("max_tokens")) { + oai_body["max_tokens"] = body.at("max_tokens"); + } else { + oai_body["max_tokens"] = 4096; + } + + // Pass through common params + for (const auto & key : {"temperature", "top_p", "top_k", "stream"}) { + if (body.contains(key)) { + oai_body[key] = body.at(key); + } + } + + // Handle Anthropic-specific thinking param + if (body.contains("thinking")) { + json thinking = json_value(body, "thinking", json::object()); + std::string thinking_type = json_value(thinking, "type", std::string()); + if (thinking_type == "enabled") { + int budget_tokens = json_value(thinking, "budget_tokens", 10000); + oai_body["thinking_budget_tokens"] = budget_tokens; + } + } + + // Handle Anthropic-specific metadata param + if (body.contains("metadata")) { + json metadata = json_value(body, "metadata", json::object()); + std::string user_id = json_value(metadata, "user_id", std::string()); + if (!user_id.empty()) { + oai_body["__metadata_user_id"] = user_id; + } + } + + return oai_body; +} + json format_embeddings_response_oaicompat(const json & request, const json & embeddings, bool use_base64) { json data = json::array(); int32_t n_tokens = 0; @@ -1211,7 +1426,7 @@ std::string tokens_to_output_formatted_string(const llama_context * ctx, const l // format server-sent event (SSE), return the formatted string to send // note: if data is a json array, it will be sent as multiple events, one per item -std::string format_sse(const json & data) { +std::string format_oai_sse(const json & data) { std::ostringstream ss; auto send_single = [&ss](const json & data) { ss << "data: " << @@ -1230,6 +1445,29 @@ std::string format_sse(const json & data) { return ss.str(); } +std::string format_anthropic_sse(const json & data) { + std::ostringstream ss; + + auto send_event = [&ss](const json & event_obj) { + if (event_obj.contains("event") && event_obj.contains("data")) { + ss << "event: " << event_obj.at("event").get() << "\n"; + ss << "data: " << safe_json_to_str(event_obj.at("data")) << "\n\n"; + } else { + ss << "data: " << safe_json_to_str(event_obj) << "\n\n"; + } + }; + + if (data.is_array()) { + for (const auto & event : data) { + send_event(event); + } + } else { + send_event(data); + } + + return ss.str(); +} + bool is_valid_utf8(const std::string & str) { const unsigned char* bytes = reinterpret_cast(str.data()); const unsigned char* end = bytes + str.length(); diff --git a/tools/server/server-common.h b/tools/server/server-common.h index 868c50610..ab8aabbad 100644 --- a/tools/server/server-common.h +++ b/tools/server/server-common.h @@ -294,6 +294,9 @@ json oaicompat_chat_params_parse( const oaicompat_parser_options & opt, std::vector & out_files); +// convert Anthropic Messages API format to OpenAI Chat Completions API format +json convert_anthropic_to_oai(const json & body); + // TODO: move it to server-task.cpp json format_embeddings_response_oaicompat(const json & request, const json & embeddings, bool use_base64 = false); @@ -320,7 +323,10 @@ std::string tokens_to_output_formatted_string(const llama_context * ctx, const l // format server-sent event (SSE), return the formatted string to send // note: if data is a json array, it will be sent as multiple events, one per item -std::string format_sse(const json & data); +std::string format_oai_sse(const json & data); + +// format Anthropic-style SSE with event types +std::string format_anthropic_sse(const json & data); bool is_valid_utf8(const std::string & str); diff --git a/tools/server/server-http.cpp b/tools/server/server-http.cpp index a82aa86b1..622505714 100644 --- a/tools/server/server-http.cpp +++ b/tools/server/server-http.cpp @@ -136,15 +136,22 @@ bool server_http_context::init(const common_params & params) { return true; } - // Check for API key in the header - auto auth_header = req.get_header_value("Authorization"); + // Check for API key in the Authorization header + std::string req_api_key = req.get_header_value("Authorization"); + if (req_api_key.empty()) { + // retry with anthropic header + req_api_key = req.get_header_value("X-Api-Key"); + } + // remove the "Bearer " prefix if needed std::string prefix = "Bearer "; - if (auth_header.substr(0, prefix.size()) == prefix) { - std::string received_api_key = auth_header.substr(prefix.size()); - if (std::find(api_keys.begin(), api_keys.end(), received_api_key) != api_keys.end()) { - return true; // API key is valid - } + if (req_api_key.substr(0, prefix.size()) == prefix) { + req_api_key = req_api_key.substr(prefix.size()); + } + + // validate the API key + if (std::find(api_keys.begin(), api_keys.end(), req_api_key) != api_keys.end()) { + return true; // API key is valid } // API key is invalid or not provided diff --git a/tools/server/server-task.cpp b/tools/server/server-task.cpp index bc4436ba6..b447a1ef6 100644 --- a/tools/server/server-task.cpp +++ b/tools/server/server-task.cpp @@ -565,15 +565,17 @@ std::vector completion_token_output::str_to_bytes(const std::stri // server_task_result_cmpl_final // json server_task_result_cmpl_final::to_json() { - switch (oaicompat) { - case OAICOMPAT_TYPE_NONE: + switch (res_type) { + case TASK_RESPONSE_TYPE_NONE: return to_json_non_oaicompat(); - case OAICOMPAT_TYPE_COMPLETION: + case TASK_RESPONSE_TYPE_OAI_CMPL: return to_json_oaicompat(); - case OAICOMPAT_TYPE_CHAT: + case TASK_RESPONSE_TYPE_OAI_CHAT: return stream ? to_json_oaicompat_chat_stream() : to_json_oaicompat_chat(); + case TASK_RESPONSE_TYPE_ANTHROPIC: + return stream ? to_json_anthropic_stream() : to_json_anthropic(); default: - GGML_ASSERT(false && "Invalid oaicompat_type"); + GGML_ASSERT(false && "Invalid task_response_type"); } } @@ -768,19 +770,203 @@ json server_task_result_cmpl_final::to_json_oaicompat_chat_stream() { return deltas; } +json server_task_result_cmpl_final::to_json_anthropic() { + std::string stop_reason = "max_tokens"; + if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) { + stop_reason = oaicompat_msg.tool_calls.empty() ? "end_turn" : "tool_use"; + } + + json content_blocks = json::array(); + + common_chat_msg msg; + if (!oaicompat_msg.empty()) { + msg = oaicompat_msg; + } else { + msg.role = "assistant"; + msg.content = content; + } + + if (!msg.content.empty()) { + content_blocks.push_back({ + {"type", "text"}, + {"text", msg.content} + }); + } + + for (const auto & tool_call : msg.tool_calls) { + json tool_use_block = { + {"type", "tool_use"}, + {"id", tool_call.id}, + {"name", tool_call.name} + }; + + try { + tool_use_block["input"] = json::parse(tool_call.arguments); + } catch (const std::exception &) { + tool_use_block["input"] = json::object(); + } + + content_blocks.push_back(tool_use_block); + } + + json res = { + {"id", oaicompat_cmpl_id}, + {"type", "message"}, + {"role", "assistant"}, + {"content", content_blocks}, + {"model", oaicompat_model}, + {"stop_reason", stop_reason}, + {"stop_sequence", stopping_word.empty() ? nullptr : json(stopping_word)}, + {"usage", { + {"input_tokens", n_prompt_tokens}, + {"output_tokens", n_decoded} + }} + }; + + return res; +} + +json server_task_result_cmpl_final::to_json_anthropic_stream() { + json events = json::array(); + + std::string stop_reason = "max_tokens"; + if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) { + stop_reason = oaicompat_msg.tool_calls.empty() ? "end_turn" : "tool_use"; + } + + bool has_text = !oaicompat_msg.content.empty(); + size_t num_tool_calls = oaicompat_msg.tool_calls.size(); + + bool text_block_started = false; + std::unordered_set tool_calls_started; + + for (const auto & diff : oaicompat_msg_diffs) { + if (!diff.content_delta.empty()) { + if (!text_block_started) { + events.push_back({ + {"event", "content_block_start"}, + {"data", { + {"type", "content_block_start"}, + {"index", 0}, + {"content_block", { + {"type", "text"}, + {"text", ""} + }} + }} + }); + text_block_started = true; + } + + events.push_back({ + {"event", "content_block_delta"}, + {"data", { + {"type", "content_block_delta"}, + {"index", 0}, + {"delta", { + {"type", "text_delta"}, + {"text", diff.content_delta} + }} + }} + }); + } + + if (diff.tool_call_index != std::string::npos) { + size_t content_block_index = (has_text ? 1 : 0) + diff.tool_call_index; + + if (tool_calls_started.find(diff.tool_call_index) == tool_calls_started.end()) { + const auto & full_tool_call = oaicompat_msg.tool_calls[diff.tool_call_index]; + + events.push_back({ + {"event", "content_block_start"}, + {"data", { + {"type", "content_block_start"}, + {"index", content_block_index}, + {"content_block", { + {"type", "tool_use"}, + {"id", full_tool_call.id}, + {"name", full_tool_call.name} + }} + }} + }); + tool_calls_started.insert(diff.tool_call_index); + } + + if (!diff.tool_call_delta.arguments.empty()) { + events.push_back({ + {"event", "content_block_delta"}, + {"data", { + {"type", "content_block_delta"}, + {"index", content_block_index}, + {"delta", { + {"type", "input_json_delta"}, + {"partial_json", diff.tool_call_delta.arguments} + }} + }} + }); + } + } + } + + if (has_text) { + events.push_back({ + {"event", "content_block_stop"}, + {"data", { + {"type", "content_block_stop"}, + {"index", 0} + }} + }); + } + + for (size_t i = 0; i < num_tool_calls; i++) { + size_t content_block_index = (has_text ? 1 : 0) + i; + events.push_back({ + {"event", "content_block_stop"}, + {"data", { + {"type", "content_block_stop"}, + {"index", content_block_index} + }} + }); + } + + events.push_back({ + {"event", "message_delta"}, + {"data", { + {"type", "message_delta"}, + {"delta", { + {"stop_reason", stop_reason}, + {"stop_sequence", stopping_word.empty() ? nullptr : json(stopping_word)} + }}, + {"usage", { + {"output_tokens", n_decoded} + }} + }} + }); + + events.push_back({ + {"event", "message_stop"}, + {"data", { + {"type", "message_stop"} + }} + }); + + return events; +} + // // server_task_result_cmpl_partial // json server_task_result_cmpl_partial::to_json() { - switch (oaicompat) { - case OAICOMPAT_TYPE_NONE: + switch (res_type) { + case TASK_RESPONSE_TYPE_NONE: return to_json_non_oaicompat(); - case OAICOMPAT_TYPE_COMPLETION: + case TASK_RESPONSE_TYPE_OAI_CMPL: return to_json_oaicompat(); - case OAICOMPAT_TYPE_CHAT: + case TASK_RESPONSE_TYPE_OAI_CHAT: return to_json_oaicompat_chat(); + case TASK_RESPONSE_TYPE_ANTHROPIC: + return to_json_anthropic(); default: - GGML_ASSERT(false && "Invalid oaicompat_type"); + GGML_ASSERT(false && "Invalid task_response_type"); } } @@ -905,7 +1091,7 @@ json server_task_result_cmpl_partial::to_json_oaicompat_chat() { // server_task_result_embd // json server_task_result_embd::to_json() { - return oaicompat == OAICOMPAT_TYPE_EMBEDDING + return res_type == TASK_RESPONSE_TYPE_OAI_EMBD ? to_json_oaicompat() : to_json_non_oaicompat(); } @@ -936,6 +1122,102 @@ json server_task_result_rerank::to_json() { }; } +json server_task_result_cmpl_partial::to_json_anthropic() { + json events = json::array(); + bool first = (n_decoded == 1); + static bool text_block_started = false; + + if (first) { + text_block_started = false; + + events.push_back({ + {"event", "message_start"}, + {"data", { + {"type", "message_start"}, + {"message", { + {"id", oaicompat_cmpl_id}, + {"type", "message"}, + {"role", "assistant"}, + {"content", json::array()}, + {"model", oaicompat_model}, + {"stop_reason", nullptr}, + {"stop_sequence", nullptr}, + {"usage", { + {"input_tokens", n_prompt_tokens}, + {"output_tokens", 0} + }} + }} + }} + }); + } + + for (const auto & diff : oaicompat_msg_diffs) { + if (!diff.content_delta.empty()) { + if (!text_block_started) { + events.push_back({ + {"event", "content_block_start"}, + {"data", { + {"type", "content_block_start"}, + {"index", 0}, + {"content_block", { + {"type", "text"}, + {"text", ""} + }} + }} + }); + text_block_started = true; + } + + events.push_back({ + {"event", "content_block_delta"}, + {"data", { + {"type", "content_block_delta"}, + {"index", 0}, + {"delta", { + {"type", "text_delta"}, + {"text", diff.content_delta} + }} + }} + }); + } + + if (diff.tool_call_index != std::string::npos) { + size_t content_block_index = (text_block_started ? 1 : 0) + diff.tool_call_index; + + if (!diff.tool_call_delta.name.empty()) { + events.push_back({ + {"event", "content_block_start"}, + {"data", { + {"type", "content_block_start"}, + {"index", content_block_index}, + {"content_block", { + {"type", "tool_use"}, + {"id", diff.tool_call_delta.id}, + {"name", diff.tool_call_delta.name} + }} + }} + }); + } + + if (!diff.tool_call_delta.arguments.empty()) { + events.push_back({ + {"event", "content_block_delta"}, + {"data", { + {"type", "content_block_delta"}, + {"index", content_block_index}, + {"delta", { + {"type", "input_json_delta"}, + {"partial_json", diff.tool_call_delta.arguments} + }} + }} + }); + } + } + } + + return events; +} + // // server_task_result_error // diff --git a/tools/server/server-task.h b/tools/server/server-task.h index 0271caae1..a22d7cab1 100644 --- a/tools/server/server-task.h +++ b/tools/server/server-task.h @@ -27,11 +27,12 @@ enum server_task_type { }; // TODO: change this to more generic "response_format" to replace the "format_response_*" in server-common -enum oaicompat_type { - OAICOMPAT_TYPE_NONE, - OAICOMPAT_TYPE_CHAT, - OAICOMPAT_TYPE_COMPLETION, - OAICOMPAT_TYPE_EMBEDDING, +enum task_response_type { + TASK_RESPONSE_TYPE_NONE, // llama.cpp native format + TASK_RESPONSE_TYPE_OAI_CHAT, + TASK_RESPONSE_TYPE_OAI_CMPL, + TASK_RESPONSE_TYPE_OAI_EMBD, + TASK_RESPONSE_TYPE_ANTHROPIC, }; enum stop_type { @@ -66,9 +67,9 @@ struct task_params { struct common_params_sampling sampling; struct common_params_speculative speculative; - // OAI-compat fields + // response formatting bool verbose = false; - oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE; + task_response_type res_type = TASK_RESPONSE_TYPE_NONE; std::string oaicompat_model; std::string oaicompat_cmpl_id; common_chat_syntax oaicompat_chat_syntax; @@ -227,12 +228,12 @@ struct server_task_result_cmpl_final : server_task_result { task_params generation_params; - // OAI-compat fields - bool verbose = false; - oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE; - std::string oaicompat_model; - std::string oaicompat_cmpl_id; - common_chat_msg oaicompat_msg; + // response formatting + bool verbose = false; + task_response_type res_type = TASK_RESPONSE_TYPE_NONE; + std::string oaicompat_model; + std::string oaicompat_cmpl_id; + common_chat_msg oaicompat_msg; std::vector oaicompat_msg_diffs; @@ -253,6 +254,10 @@ struct server_task_result_cmpl_final : server_task_result { json to_json_oaicompat_chat(); json to_json_oaicompat_chat_stream(); + + json to_json_anthropic(); + + json to_json_anthropic_stream(); }; struct server_task_result_cmpl_partial : server_task_result { @@ -270,11 +275,11 @@ struct server_task_result_cmpl_partial : server_task_result { result_timings timings; result_prompt_progress progress; - // OAI-compat fields - bool verbose = false; - oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE; - std::string oaicompat_model; - std::string oaicompat_cmpl_id; + // response formatting + bool verbose = false; + task_response_type res_type = TASK_RESPONSE_TYPE_NONE; + std::string oaicompat_model; + std::string oaicompat_cmpl_id; std::vector oaicompat_msg_diffs; virtual int get_index() override { @@ -292,6 +297,8 @@ struct server_task_result_cmpl_partial : server_task_result { json to_json_oaicompat(); json to_json_oaicompat_chat(); + + json to_json_anthropic(); }; struct server_task_result_embd : server_task_result { @@ -300,8 +307,8 @@ struct server_task_result_embd : server_task_result { int32_t n_tokens; - // OAI-compat fields - oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE; + // response formatting + task_response_type res_type = TASK_RESPONSE_TYPE_NONE; virtual int get_index() override { return index; diff --git a/tools/server/server.cpp b/tools/server/server.cpp index 0f39def37..05bbe648c 100644 --- a/tools/server/server.cpp +++ b/tools/server/server.cpp @@ -1255,7 +1255,7 @@ struct server_context { res->post_sampling_probs = slot.task->params.post_sampling_probs; res->verbose = slot.task->params.verbose; - res->oaicompat = slot.task->params.oaicompat; + res->res_type = slot.task->params.res_type; res->oaicompat_model = slot.task->params.oaicompat_model; res->oaicompat_cmpl_id = slot.task->params.oaicompat_cmpl_id; @@ -1297,7 +1297,7 @@ struct server_context { res->verbose = slot.task->params.verbose; res->stream = slot.task->params.stream; res->include_usage = slot.task->params.include_usage; - res->oaicompat = slot.task->params.oaicompat; + res->res_type = slot.task->params.res_type; res->oaicompat_model = slot.task->params.oaicompat_model; res->oaicompat_cmpl_id = slot.task->params.oaicompat_cmpl_id; res->oaicompat_msg = slot.update_chat_msg(res->oaicompat_msg_diffs); @@ -1328,7 +1328,7 @@ struct server_context { res->id = slot.task->id; res->index = slot.task->index; res->n_tokens = slot.task->n_tokens(); - res->oaicompat = slot.task->params.oaicompat; + res->res_type = slot.task->params.res_type; const int n_embd = llama_model_n_embd(model); @@ -2951,7 +2951,7 @@ public: data, files, req.should_stop, - OAICOMPAT_TYPE_NONE); // infill is not OAI compatible + TASK_RESPONSE_TYPE_NONE); // infill is not OAI compatible }; server_http_context::handler_t post_completions = [this](const server_http_req & req) { @@ -2962,7 +2962,7 @@ public: body, files, req.should_stop, - OAICOMPAT_TYPE_NONE); + TASK_RESPONSE_TYPE_NONE); }; server_http_context::handler_t post_completions_oai = [this](const server_http_req & req) { @@ -2973,7 +2973,7 @@ public: body, files, req.should_stop, - OAICOMPAT_TYPE_COMPLETION); + TASK_RESPONSE_TYPE_OAI_CMPL); }; server_http_context::handler_t post_chat_completions = [this](const server_http_req & req) { @@ -2988,7 +2988,38 @@ public: body_parsed, files, req.should_stop, - OAICOMPAT_TYPE_CHAT); + TASK_RESPONSE_TYPE_OAI_CHAT); + }; + + server_http_context::handler_t post_anthropic_messages = [this](const server_http_req & req) { + std::vector files; + json body = convert_anthropic_to_oai(json::parse(req.body)); + json body_parsed = oaicompat_chat_params_parse( + body, + ctx_server.oai_parser_opt, + files); + return handle_completions_impl( + SERVER_TASK_TYPE_COMPLETION, + body_parsed, + files, + req.should_stop, + TASK_RESPONSE_TYPE_ANTHROPIC); + }; + + server_http_context::handler_t post_anthropic_count_tokens = [this](const server_http_req & req) { + auto res = std::make_unique(ctx_server); + std::vector files; + json body = convert_anthropic_to_oai(json::parse(req.body)); + json body_parsed = oaicompat_chat_params_parse( + body, + ctx_server.oai_parser_opt, + files); + + json prompt = body_parsed.at("prompt"); + llama_tokens tokens = tokenize_mixed(ctx_server.vocab, prompt, true, true); + + res->ok({{"input_tokens", static_cast(tokens.size())}}); + return res; }; // same with handle_chat_completions, but without inference part @@ -3107,11 +3138,11 @@ public: }; server_http_context::handler_t post_embeddings = [this](const server_http_req & req) { - return handle_embeddings_impl(req, OAICOMPAT_TYPE_NONE); + return handle_embeddings_impl(req, TASK_RESPONSE_TYPE_NONE); }; server_http_context::handler_t post_embeddings_oai = [this](const server_http_req & req) { - return handle_embeddings_impl(req, OAICOMPAT_TYPE_EMBEDDING); + return handle_embeddings_impl(req, TASK_RESPONSE_TYPE_OAI_EMBD); }; server_http_context::handler_t post_rerank = [this](const server_http_req & req) { @@ -3262,7 +3293,7 @@ private: const json & data, const std::vector & files, const std::function & should_stop, - oaicompat_type oaicompat) { + task_response_type res_type) { GGML_ASSERT(type == SERVER_TASK_TYPE_COMPLETION || type == SERVER_TASK_TYPE_INFILL); auto res = std::make_unique(ctx_server); @@ -3279,7 +3310,7 @@ private: // process prompt std::vector inputs; - if (oaicompat && ctx_server.mctx != nullptr) { + if (res_type != TASK_RESPONSE_TYPE_NONE && ctx_server.mctx != nullptr) { // This is the case used by OAI compatible chat path with MTMD. TODO It can be moved to the path below. inputs.push_back(process_mtmd_prompt(ctx_server.mctx, prompt.get(), files)); } else { @@ -3301,8 +3332,8 @@ private: task.id_slot = json_value(data, "id_slot", -1); // OAI-compat - task.params.oaicompat = oaicompat; - task.params.oaicompat_cmpl_id = completion_id; + task.params.res_type = res_type; + task.params.oaicompat_cmpl_id = completion_id; // oaicompat_model is already populated by params_from_json_cmpl tasks.push_back(std::move(task)); @@ -3352,10 +3383,14 @@ private: } // next responses are streamed - res->data = format_sse(first_result->to_json()); // to be sent immediately + if (res_type == TASK_RESPONSE_TYPE_ANTHROPIC) { + res->data = format_anthropic_sse(first_result->to_json()); + } else { + res->data = format_oai_sse(first_result->to_json()); // to be sent immediately + } res->status = 200; res->content_type = "text/event-stream"; - res->next = [res_this = res.get(), oaicompat, &should_stop](std::string & output) -> bool { + res->next = [res_this = res.get(), res_type, &should_stop](std::string & output) -> bool { if (should_stop()) { SRV_DBG("%s", "stopping streaming due to should_stop condition\n"); return false; // should_stop condition met @@ -3372,7 +3407,10 @@ private: // check if there is more data if (!rd.has_next()) { - if (oaicompat != OAICOMPAT_TYPE_NONE) { + if (res_type == TASK_RESPONSE_TYPE_ANTHROPIC) { + // Anthropic doesn't send [DONE], message_stop was already sent + output = ""; + } else if (res_type != TASK_RESPONSE_TYPE_NONE) { output = "data: [DONE]\n\n"; } else { output = ""; @@ -3391,7 +3429,14 @@ private: // send the results json res_json = result->to_json(); if (result->is_error()) { - output = format_sse(json {{ "error", res_json }}); + if (res_type == TASK_RESPONSE_TYPE_ANTHROPIC) { + output = format_anthropic_sse({ + {"event", "error"}, + {"data", res_json}, + }); + } else { + output = format_oai_sse(json {{ "error", res_json }}); + } SRV_DBG("%s", "error received during streaming, terminating stream\n"); return false; // terminate on error } else { @@ -3399,7 +3444,11 @@ private: dynamic_cast(result.get()) != nullptr || dynamic_cast(result.get()) != nullptr ); - output = format_sse(res_json); + if (res_type == TASK_RESPONSE_TYPE_ANTHROPIC) { + output = format_anthropic_sse(res_json); + } else { + output = format_oai_sse(res_json); + } } // has next data, continue @@ -3507,14 +3556,14 @@ private: return res; } - std::unique_ptr handle_embeddings_impl(const server_http_req & req, oaicompat_type oaicompat) { + std::unique_ptr handle_embeddings_impl(const server_http_req & req, task_response_type res_type) { auto res = std::make_unique(ctx_server); if (!ctx_server.params_base.embedding) { res->error(format_error_response("This server does not support embeddings. Start it with `--embeddings`", ERROR_TYPE_NOT_SUPPORTED)); return res; } - if (oaicompat != OAICOMPAT_TYPE_NONE && llama_pooling_type(ctx_server.ctx) == LLAMA_POOLING_TYPE_NONE) { + if (res_type != TASK_RESPONSE_TYPE_NONE && llama_pooling_type(ctx_server.ctx) == LLAMA_POOLING_TYPE_NONE) { res->error(format_error_response("Pooling type 'none' is not OAI compatible. Please use a different pooling type", ERROR_TYPE_INVALID_REQUEST)); return res; } @@ -3526,7 +3575,7 @@ private: if (body.count("input") != 0) { prompt = body.at("input"); } else if (body.contains("content")) { - oaicompat = OAICOMPAT_TYPE_NONE; // "content" field is not OAI compatible + res_type = TASK_RESPONSE_TYPE_NONE; // "content" field is not OAI compatible prompt = body.at("content"); } else { res->error(format_error_response("\"input\" or \"content\" must be provided", ERROR_TYPE_INVALID_REQUEST)); @@ -3574,7 +3623,7 @@ private: task.tokens = std::move(tokenized_prompts[i]); // OAI-compat - task.params.oaicompat = oaicompat; + task.params.res_type = res_type; task.params.embd_normalize = embd_normalize; tasks.push_back(std::move(task)); @@ -3599,7 +3648,7 @@ private: } // write JSON response - json root = oaicompat == OAICOMPAT_TYPE_EMBEDDING + json root = res_type == TASK_RESPONSE_TYPE_OAI_EMBD ? format_embeddings_response_oaicompat(body, responses, use_base64) : json(responses); res->ok(root); @@ -3712,6 +3761,8 @@ int main(int argc, char ** argv) { ctx_http.post("/chat/completions", ex_wrapper(routes.post_chat_completions)); ctx_http.post("/v1/chat/completions", ex_wrapper(routes.post_chat_completions)); ctx_http.post("/api/chat", ex_wrapper(routes.post_chat_completions)); // ollama specific endpoint + ctx_http.post("/v1/messages", ex_wrapper(routes.post_anthropic_messages)); // anthropic messages API + ctx_http.post("/v1/messages/count_tokens", ex_wrapper(routes.post_anthropic_count_tokens)); // anthropic token counting ctx_http.post("/infill", ex_wrapper(routes.post_infill)); ctx_http.post("/embedding", ex_wrapper(routes.post_embeddings)); // legacy ctx_http.post("/embeddings", ex_wrapper(routes.post_embeddings)); diff --git a/tools/server/tests/conftest.py b/tools/server/tests/conftest.py index 017d1bb84..c7ed77596 100644 --- a/tools/server/tests/conftest.py +++ b/tools/server/tests/conftest.py @@ -13,3 +13,9 @@ def stop_server_after_each_test(): ) # copy the set to prevent 'Set changed size during iteration' for server in instances: server.stop() + + +@pytest.fixture(scope="module", autouse=True) +def do_something(): + # this will be run once per test session, before any tests + ServerPreset.load_all() diff --git a/tools/server/tests/unit/test_basic.py b/tools/server/tests/unit/test_basic.py index 720b136b0..cadaa9184 100644 --- a/tools/server/tests/unit/test_basic.py +++ b/tools/server/tests/unit/test_basic.py @@ -5,12 +5,6 @@ from utils import * server = ServerPreset.tinyllama2() -@pytest.fixture(scope="session", autouse=True) -def do_something(): - # this will be run once per test session, before any tests - ServerPreset.load_all() - - @pytest.fixture(autouse=True) def create_server(): global server diff --git a/tools/server/tests/unit/test_compat_anthropic.py b/tools/server/tests/unit/test_compat_anthropic.py new file mode 100644 index 000000000..d55dd1d94 --- /dev/null +++ b/tools/server/tests/unit/test_compat_anthropic.py @@ -0,0 +1,807 @@ +#!/usr/bin/env python3 +import pytest +import base64 +import requests + +from utils import * + +server: ServerProcess + + +def get_test_image_base64() -> str: + """Get a test image in base64 format""" + # Use the same test image as test_vision_api.py + IMG_URL = "https://huggingface.co/ggml-org/tinygemma3-GGUF/resolve/main/test/11_truck.png" + response = requests.get(IMG_URL) + response.raise_for_status() + return base64.b64encode(response.content).decode("utf-8") + +@pytest.fixture(autouse=True) +def create_server(): + global server + server = ServerPreset.tinyllama2() + server.model_alias = "tinyllama-2-anthropic" + server.server_port = 8082 + server.n_slots = 1 + server.n_ctx = 8192 + server.n_batch = 2048 + + +@pytest.fixture +def vision_server(): + """Separate fixture for vision tests that require multimodal support""" + global server + server = ServerPreset.tinygemma3() + server.offline = False # Allow downloading the model + server.model_alias = "tinygemma3-anthropic" + server.server_port = 8083 # Different port to avoid conflicts + server.n_slots = 1 + return server + + +# Basic message tests + +def test_anthropic_messages_basic(): + """Test basic Anthropic messages endpoint""" + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 50, + "messages": [ + {"role": "user", "content": "Say hello"} + ] + }) + + assert res.status_code == 200, f"Expected 200, got {res.status_code}" + assert res.body["type"] == "message", f"Expected type 'message', got {res.body.get('type')}" + assert res.body["role"] == "assistant", f"Expected role 'assistant', got {res.body.get('role')}" + assert "content" in res.body, "Missing 'content' field" + assert isinstance(res.body["content"], list), "Content should be an array" + assert len(res.body["content"]) > 0, "Content array should not be empty" + assert res.body["content"][0]["type"] == "text", "First content block should be text" + assert "text" in res.body["content"][0], "Text content block missing 'text' field" + assert res.body["stop_reason"] in ["end_turn", "max_tokens"], f"Invalid stop_reason: {res.body.get('stop_reason')}" + assert "usage" in res.body, "Missing 'usage' field" + assert "input_tokens" in res.body["usage"], "Missing usage.input_tokens" + assert "output_tokens" in res.body["usage"], "Missing usage.output_tokens" + assert isinstance(res.body["usage"]["input_tokens"], int), "input_tokens should be integer" + assert isinstance(res.body["usage"]["output_tokens"], int), "output_tokens should be integer" + assert res.body["usage"]["output_tokens"] > 0, "Should have generated some tokens" + # Anthropic API should NOT include timings + assert "timings" not in res.body, "Anthropic API should not include timings field" + + +def test_anthropic_messages_with_system(): + """Test messages with system prompt""" + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 50, + "system": "You are a helpful assistant.", + "messages": [ + {"role": "user", "content": "Hello"} + ] + }) + + assert res.status_code == 200 + assert res.body["type"] == "message" + assert len(res.body["content"]) > 0 + + +def test_anthropic_messages_multipart_content(): + """Test messages with multipart content blocks""" + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 50, + "messages": [ + { + "role": "user", + "content": [ + {"type": "text", "text": "What is"}, + {"type": "text", "text": " the answer?"} + ] + } + ] + }) + + assert res.status_code == 200 + assert res.body["type"] == "message" + + +def test_anthropic_messages_conversation(): + """Test multi-turn conversation""" + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 50, + "messages": [ + {"role": "user", "content": "Hello"}, + {"role": "assistant", "content": "Hi there!"}, + {"role": "user", "content": "How are you?"} + ] + }) + + assert res.status_code == 200 + assert res.body["type"] == "message" + + +# Streaming tests + +def test_anthropic_messages_streaming(): + """Test streaming messages""" + server.start() + + res = server.make_stream_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 30, + "messages": [ + {"role": "user", "content": "Say hello"} + ], + "stream": True + }) + + events = [] + for data in res: + # Each event should have type and other fields + assert "type" in data, f"Missing 'type' in event: {data}" + events.append(data) + + # Verify event sequence + event_types = [e["type"] for e in events] + assert "message_start" in event_types, "Missing message_start event" + assert "content_block_start" in event_types, "Missing content_block_start event" + assert "content_block_delta" in event_types, "Missing content_block_delta event" + assert "content_block_stop" in event_types, "Missing content_block_stop event" + assert "message_delta" in event_types, "Missing message_delta event" + assert "message_stop" in event_types, "Missing message_stop event" + + # Check message_start structure + message_start = next(e for e in events if e["type"] == "message_start") + assert "message" in message_start, "message_start missing 'message' field" + assert message_start["message"]["type"] == "message" + assert message_start["message"]["role"] == "assistant" + assert message_start["message"]["content"] == [] + assert "usage" in message_start["message"] + assert message_start["message"]["usage"]["input_tokens"] > 0 + + # Check content_block_start + block_start = next(e for e in events if e["type"] == "content_block_start") + assert "index" in block_start, "content_block_start missing 'index'" + assert block_start["index"] == 0, "First content block should be at index 0" + assert "content_block" in block_start + assert block_start["content_block"]["type"] == "text" + + # Check content_block_delta + deltas = [e for e in events if e["type"] == "content_block_delta"] + assert len(deltas) > 0, "Should have at least one content_block_delta" + for delta in deltas: + assert "index" in delta + assert "delta" in delta + assert delta["delta"]["type"] == "text_delta" + assert "text" in delta["delta"] + + # Check content_block_stop + block_stop = next(e for e in events if e["type"] == "content_block_stop") + assert "index" in block_stop + assert block_stop["index"] == 0 + + # Check message_delta + message_delta = next(e for e in events if e["type"] == "message_delta") + assert "delta" in message_delta + assert "stop_reason" in message_delta["delta"] + assert message_delta["delta"]["stop_reason"] in ["end_turn", "max_tokens"] + assert "usage" in message_delta + assert message_delta["usage"]["output_tokens"] > 0 + + # Check message_stop + message_stop = next(e for e in events if e["type"] == "message_stop") + # message_stop should NOT have timings for Anthropic API + assert "timings" not in message_stop, "Anthropic streaming should not include timings" + + +# Token counting tests + +def test_anthropic_count_tokens(): + """Test token counting endpoint""" + server.start() + + res = server.make_request("POST", "/v1/messages/count_tokens", data={ + "model": "test", + "messages": [ + {"role": "user", "content": "Hello world"} + ] + }) + + assert res.status_code == 200 + assert "input_tokens" in res.body + assert isinstance(res.body["input_tokens"], int) + assert res.body["input_tokens"] > 0 + # Should only have input_tokens, no other fields + assert "output_tokens" not in res.body + + +def test_anthropic_count_tokens_with_system(): + """Test token counting with system prompt""" + server.start() + + res = server.make_request("POST", "/v1/messages/count_tokens", data={ + "model": "test", + "system": "You are a helpful assistant.", + "messages": [ + {"role": "user", "content": "Hello"} + ] + }) + + assert res.status_code == 200 + assert res.body["input_tokens"] > 0 + + +def test_anthropic_count_tokens_no_max_tokens(): + """Test that count_tokens doesn't require max_tokens""" + server.start() + + # max_tokens is NOT required for count_tokens + res = server.make_request("POST", "/v1/messages/count_tokens", data={ + "model": "test", + "messages": [ + {"role": "user", "content": "Hello"} + ] + }) + + assert res.status_code == 200 + assert "input_tokens" in res.body + + +# Tool use tests + +def test_anthropic_tool_use_basic(): + """Test basic tool use""" + server.jinja = True + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 200, + "tools": [{ + "name": "get_weather", + "description": "Get the current weather in a location", + "input_schema": { + "type": "object", + "properties": { + "location": { + "type": "string", + "description": "City name" + } + }, + "required": ["location"] + } + }], + "messages": [ + {"role": "user", "content": "What's the weather in Paris?"} + ] + }) + + assert res.status_code == 200 + assert res.body["type"] == "message" + assert len(res.body["content"]) > 0 + + # Check if model used the tool (it might not always, depending on the model) + content_types = [block.get("type") for block in res.body["content"]] + + if "tool_use" in content_types: + # Model used the tool + assert res.body["stop_reason"] == "tool_use" + + # Find the tool_use block + tool_block = next(b for b in res.body["content"] if b.get("type") == "tool_use") + assert "id" in tool_block + assert "name" in tool_block + assert tool_block["name"] == "get_weather" + assert "input" in tool_block + assert isinstance(tool_block["input"], dict) + + +def test_anthropic_tool_result(): + """Test sending tool results back + + This test verifies that tool_result blocks are properly converted to + role="tool" messages internally. Without proper conversion, this would + fail with a 500 error: "unsupported content[].type" because tool_result + blocks would remain in the user message content array. + """ + server.jinja = True + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 100, + "messages": [ + {"role": "user", "content": "What's the weather?"}, + { + "role": "assistant", + "content": [ + { + "type": "tool_use", + "id": "test123", + "name": "get_weather", + "input": {"location": "Paris"} + } + ] + }, + { + "role": "user", + "content": [ + { + "type": "tool_result", + "tool_use_id": "test123", + "content": "The weather is sunny, 25°C" + } + ] + } + ] + }) + + # This would be 500 with the old bug where tool_result blocks weren't converted + assert res.status_code == 200 + assert res.body["type"] == "message" + # Model should respond to the tool result + assert len(res.body["content"]) > 0 + assert res.body["content"][0]["type"] == "text" + + +def test_anthropic_tool_result_with_text(): + """Test tool result mixed with text content + + This tests the edge case where a user message contains both text and + tool_result blocks. The server must properly split these into separate + messages: a user message with text, followed by tool messages. + Without proper handling, this would fail with 500: "unsupported content[].type" + """ + server.jinja = True + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 100, + "messages": [ + {"role": "user", "content": "What's the weather?"}, + { + "role": "assistant", + "content": [ + { + "type": "tool_use", + "id": "tool_1", + "name": "get_weather", + "input": {"location": "Paris"} + } + ] + }, + { + "role": "user", + "content": [ + {"type": "text", "text": "Here are the results:"}, + { + "type": "tool_result", + "tool_use_id": "tool_1", + "content": "Sunny, 25°C" + } + ] + } + ] + }) + + assert res.status_code == 200 + assert res.body["type"] == "message" + assert len(res.body["content"]) > 0 + + +def test_anthropic_tool_result_error(): + """Test tool result with error flag""" + server.jinja = True + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 100, + "messages": [ + {"role": "user", "content": "Get the weather"}, + { + "role": "assistant", + "content": [ + { + "type": "tool_use", + "id": "test123", + "name": "get_weather", + "input": {"location": "InvalidCity"} + } + ] + }, + { + "role": "user", + "content": [ + { + "type": "tool_result", + "tool_use_id": "test123", + "is_error": True, + "content": "City not found" + } + ] + } + ] + }) + + assert res.status_code == 200 + assert res.body["type"] == "message" + + +def test_anthropic_tool_streaming(): + """Test streaming with tool use""" + server.jinja = True + server.start() + + res = server.make_stream_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 200, + "stream": True, + "tools": [{ + "name": "calculator", + "description": "Calculate math", + "input_schema": { + "type": "object", + "properties": { + "expression": {"type": "string"} + }, + "required": ["expression"] + } + }], + "messages": [ + {"role": "user", "content": "Calculate 2+2"} + ] + }) + + events = [] + for data in res: + events.append(data) + + event_types = [e["type"] for e in events] + + # Should have basic events + assert "message_start" in event_types + assert "message_stop" in event_types + + # If tool was used, check for proper tool streaming + if any(e.get("type") == "content_block_start" and + e.get("content_block", {}).get("type") == "tool_use" + for e in events): + # Find tool use block start + tool_starts = [e for e in events if + e.get("type") == "content_block_start" and + e.get("content_block", {}).get("type") == "tool_use"] + + assert len(tool_starts) > 0, "Should have tool_use content_block_start" + + # Check index is correct (should be 0 if no text, 1 if there's text) + tool_start = tool_starts[0] + assert "index" in tool_start + assert tool_start["content_block"]["type"] == "tool_use" + assert "name" in tool_start["content_block"] + + +# Vision/multimodal tests + +def test_anthropic_vision_format_accepted(): + """Test that Anthropic vision format is accepted (format validation only)""" + server.start() + + # Small 1x1 red PNG image in base64 + red_pixel_png = "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mP8z8DwHwAFBQIAX8jx0gAAAABJRU5ErkJggg==" + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 10, + "messages": [ + { + "role": "user", + "content": [ + { + "type": "image", + "source": { + "type": "base64", + "media_type": "image/png", + "data": red_pixel_png + } + }, + { + "type": "text", + "text": "What is this?" + } + ] + } + ] + }) + + # Server accepts the format but tinyllama doesn't support images + # So it should return 500 with clear error message about missing mmproj + assert res.status_code == 500 + assert "image input is not supported" in res.body.get("error", {}).get("message", "").lower() + + +def test_anthropic_vision_base64_with_multimodal_model(vision_server): + """Test vision with base64 image using Anthropic format with multimodal model""" + global server + server = vision_server + server.start() + + # Get test image in base64 format + image_base64 = get_test_image_base64() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 10, + "messages": [ + { + "role": "user", + "content": [ + { + "type": "image", + "source": { + "type": "base64", + "media_type": "image/png", + "data": image_base64 + } + }, + { + "type": "text", + "text": "What is this:\n" + } + ] + } + ] + }) + + assert res.status_code == 200, f"Expected 200, got {res.status_code}: {res.body}" + assert res.body["type"] == "message" + assert len(res.body["content"]) > 0 + assert res.body["content"][0]["type"] == "text" + # The model should generate some response about the image + assert len(res.body["content"][0]["text"]) > 0 + + +# Parameter tests + +def test_anthropic_stop_sequences(): + """Test stop_sequences parameter""" + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 100, + "stop_sequences": ["\n", "END"], + "messages": [ + {"role": "user", "content": "Count to 10"} + ] + }) + + assert res.status_code == 200 + assert res.body["type"] == "message" + + +def test_anthropic_temperature(): + """Test temperature parameter""" + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 50, + "temperature": 0.5, + "messages": [ + {"role": "user", "content": "Hello"} + ] + }) + + assert res.status_code == 200 + assert res.body["type"] == "message" + + +def test_anthropic_top_p(): + """Test top_p parameter""" + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 50, + "top_p": 0.9, + "messages": [ + {"role": "user", "content": "Hello"} + ] + }) + + assert res.status_code == 200 + assert res.body["type"] == "message" + + +def test_anthropic_top_k(): + """Test top_k parameter (llama.cpp specific)""" + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 50, + "top_k": 40, + "messages": [ + {"role": "user", "content": "Hello"} + ] + }) + + assert res.status_code == 200 + assert res.body["type"] == "message" + + +# Error handling tests + +def test_anthropic_missing_messages(): + """Test error when messages are missing""" + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 50 + # missing "messages" field + }) + + # Should return an error (400 or 500) + assert res.status_code >= 400 + + +def test_anthropic_empty_messages(): + """Test permissive handling of empty messages array""" + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 50, + "messages": [] + }) + + # Server is permissive and accepts empty messages (provides defaults) + # This matches the permissive validation design choice + assert res.status_code == 200 + assert res.body["type"] == "message" + + +# Content block index tests + +def test_anthropic_streaming_content_block_indices(): + """Test that content block indices are correct in streaming""" + server.jinja = True + server.start() + + # Request that might produce both text and tool use + res = server.make_stream_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 200, + "stream": True, + "tools": [{ + "name": "test_tool", + "description": "A test tool", + "input_schema": { + "type": "object", + "properties": { + "param": {"type": "string"} + }, + "required": ["param"] + } + }], + "messages": [ + {"role": "user", "content": "Use the test tool"} + ] + }) + + events = [] + for data in res: + events.append(data) + + # Check content_block_start events have sequential indices + block_starts = [e for e in events if e.get("type") == "content_block_start"] + if len(block_starts) > 1: + # If there are multiple blocks, indices should be sequential + indices = [e["index"] for e in block_starts] + expected_indices = list(range(len(block_starts))) + assert indices == expected_indices, f"Expected indices {expected_indices}, got {indices}" + + # Check content_block_stop events match the starts + block_stops = [e for e in events if e.get("type") == "content_block_stop"] + start_indices = set(e["index"] for e in block_starts) + stop_indices = set(e["index"] for e in block_stops) + assert start_indices == stop_indices, "content_block_stop indices should match content_block_start indices" + + +# Extended features tests + +def test_anthropic_thinking(): + """Test extended thinking parameter""" + server.jinja = True + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 100, + "thinking": { + "type": "enabled", + "budget_tokens": 50 + }, + "messages": [ + {"role": "user", "content": "What is 2+2?"} + ] + }) + + assert res.status_code == 200 + assert res.body["type"] == "message" + + +def test_anthropic_metadata(): + """Test metadata parameter""" + server.start() + + res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 50, + "metadata": { + "user_id": "test_user_123" + }, + "messages": [ + {"role": "user", "content": "Hello"} + ] + }) + + assert res.status_code == 200 + assert res.body["type"] == "message" + + +# Compatibility tests + +def test_anthropic_vs_openai_different_response_format(): + """Verify Anthropic format is different from OpenAI format""" + server.start() + + # Make OpenAI request + openai_res = server.make_request("POST", "/v1/chat/completions", data={ + "model": "test", + "max_tokens": 50, + "messages": [ + {"role": "user", "content": "Hello"} + ] + }) + + # Make Anthropic request + anthropic_res = server.make_request("POST", "/v1/messages", data={ + "model": "test", + "max_tokens": 50, + "messages": [ + {"role": "user", "content": "Hello"} + ] + }) + + assert openai_res.status_code == 200 + assert anthropic_res.status_code == 200 + + # OpenAI has "object", Anthropic has "type" + assert "object" in openai_res.body + assert "type" in anthropic_res.body + assert openai_res.body["object"] == "chat.completion" + assert anthropic_res.body["type"] == "message" + + # OpenAI has "choices", Anthropic has "content" + assert "choices" in openai_res.body + assert "content" in anthropic_res.body + + # Different usage field names + assert "prompt_tokens" in openai_res.body["usage"] + assert "input_tokens" in anthropic_res.body["usage"] + assert "completion_tokens" in openai_res.body["usage"] + assert "output_tokens" in anthropic_res.body["usage"] diff --git a/tools/server/tests/unit/test_security.py b/tools/server/tests/unit/test_security.py index 0e1158055..e160a8e6d 100644 --- a/tools/server/tests/unit/test_security.py +++ b/tools/server/tests/unit/test_security.py @@ -49,6 +49,19 @@ def test_correct_api_key(): assert "content" in res.body +def test_correct_api_key_anthropic_header(): + global server + server.start() + res = server.make_request("POST", "/completions", data={ + "prompt": "I believe the meaning of life is", + }, headers={ + "X-Api-Key": TEST_API_KEY, + }) + assert res.status_code == 200 + assert "error" not in res.body + assert "content" in res.body + + def test_openai_library_correct_api_key(): global server server.start()