diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index f9baf687e..0c9872228 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -684,6 +684,7 @@ class MODEL_ARCH(IntEnum): SPARK2_5 = auto() LLAMA_EMBED = auto() MAINCODER = auto() + MAION_CODER = auto() KIMI_LINEAR = auto() KIMI_K3 = auto() TALKIE = auto() @@ -1488,6 +1489,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.SPARK2_5: "spark2_5", MODEL_ARCH.LLAMA_EMBED: "llama-embed", MODEL_ARCH.MAINCODER: "maincoder", + MODEL_ARCH.MAION_CODER: "maion-coder", MODEL_ARCH.KIMI_LINEAR: "kimi-linear", MODEL_ARCH.KIMI_K3: "kimi-k3", MODEL_ARCH.TALKIE: "talkie", @@ -5742,6 +5744,31 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.MAION_CODER: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + 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.FFN_NORM, + MODEL_TENSOR.FFN_POST_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.MOE_LATENT_DOWN, + MODEL_TENSOR.MOE_LATENT_UP, + ], MODEL_ARCH.KIMI_LINEAR: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 5744c5fb8..d1dd4cbae 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -154,6 +154,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_SPARK2_5, "spark2_5" }, { LLM_ARCH_LLAMA_EMBED, "llama-embed" }, { LLM_ARCH_MAINCODER, "maincoder" }, + { LLM_ARCH_MAION_CODER, "maion-coder" }, { LLM_ARCH_KIMI_LINEAR, "kimi-linear" }, { LLM_ARCH_KIMI_K3, "kimi-k3" }, { LLM_ARCH_GLM5_NEXT, "glm5-next" }, diff --git a/src/llama-arch.h b/src/llama-arch.h index 94771c2a1..4e8e53421 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -154,6 +154,7 @@ enum llm_arch { LLM_ARCH_SPARK2_5, LLM_ARCH_LLAMA_EMBED, LLM_ARCH_MAINCODER, + LLM_ARCH_MAION_CODER, LLM_ARCH_KIMI_LINEAR, LLM_ARCH_KIMI_K3, LLM_ARCH_GLM5_NEXT, diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp index 60a7e8554..b7dfadef4 100644 --- a/src/llama-kv-cache.cpp +++ b/src/llama-kv-cache.cpp @@ -2009,7 +2009,7 @@ public: void llm_graph_input_k_shift::set_input(const llama_ubatch * ubatch) { GGML_UNUSED(ubatch); - if (k_shift) { + if (k_shift && k_shift->buffer) { kv_self->set_input_k_shift(k_shift); } diff --git a/src/llama-model-saver.cpp b/src/llama-model-saver.cpp index 47e36d374..e011e24ea 100644 --- a/src/llama-model-saver.cpp +++ b/src/llama-model-saver.cpp @@ -225,8 +225,9 @@ void llama_model_saver::add_kv_from_model() { add_kv(LLM_KV_BLOCK_COUNT, hparams.n_layer_all); add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead); add_kv(LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff_arr, true); - add_kv(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp()); + add_kv(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, true); add_kv(LLM_KV_EXPERT_LATENT_LENGTH, hparams.n_expert_latent); + add_kv(LLM_KV_MOE_LATENT_SIZE, hparams.moe_latent_size); add_kv(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp); add_kv(LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH, hparams.n_ff_chexp); add_kv(LLM_KV_SWIGLU_CLAMP_EXP, std::vector( diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 923c2c0a2..ded790b2b 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -52,6 +52,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params return new llama_model_llama_embed(params); case LLM_ARCH_MAINCODER: return new llama_model_maincoder(params); + case LLM_ARCH_MAION_CODER: + return new llama_model_maion_coder(params); case LLM_ARCH_TALKIE: return new llama_model_talkie(params); case LLM_ARCH_DECI: @@ -3556,6 +3558,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_GEMMA3: case LLM_ARCH_GEMMA3N: case LLM_ARCH_GEMMA4: + case LLM_ARCH_MAION_CODER: case LLM_ARCH_GEMMA4_ASSISTANT: case LLM_ARCH_GEMMA_EMBEDDING: case LLM_ARCH_GEMMA_EMBEDDING2: diff --git a/src/models/maion-coder.cpp b/src/models/maion-coder.cpp new file mode 100644 index 000000000..bb3ebabc7 --- /dev/null +++ b/src/models/maion-coder.cpp @@ -0,0 +1,149 @@ +#include "models.h" + +void llama_model_maion_coder::load_arch_hparams(llama_model_loader & ml) { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all); + ml.get_key(LLM_KV_INTERLEAVE_MOE_LAYER_STEP, hparams.n_moe_layer_step); + ml.get_key(LLM_KV_MOE_LATENT_SIZE, hparams.moe_latent_size); + + if (hparams.n_moe_layer_step == 0 || hparams.moe_latent_size == 0 || hparams.n_expert == 0) { + throw std::runtime_error("maion-coder requires a positive MoE layer step, latent size and expert count"); + } + + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa); + uint32_t n_swa_layers = 0; + ml.get_arr_n(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, n_swa_layers); + if (hparams.n_swa == 0 || n_swa_layers != hparams.n_layer_all) { + throw std::runtime_error("maion-coder requires a positive sliding window and a sliding window pattern for each layer"); + } + ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl); + hparams.rope_pattern = hparams.is_swa_impl; + hparams.swa_type = LLAMA_SWA_TYPE_STANDARD; + hparams.rope_freq_base_train_swa = hparams.rope_freq_base_train; + hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train; + + type = LLM_TYPE_UNKNOWN; +} + +void llama_model_maion_coder::load_arch_tensors(llama_model_loader &) { + LLAMA_LOAD_LOCALS; + + const int64_t moe_n_embd = hparams.moe_latent_size; + + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + 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}, 0); + + for (int i = 0; i < n_layer; ++i) { + const bool is_moe_layer = (i + 1) % hparams.n_moe_layer_step == 0; + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, 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); + 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); + layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0); + + if (is_moe_layer) { + const int64_t n_ff_exp = hparams.n_ff_exp(i); + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0); + layer.ffn_latent_down = create_tensor(tn(LLM_TENSOR_FFN_LATENT_DOWN, "weight", i), {n_embd, moe_n_embd}, 0); + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {moe_n_embd, n_ff_exp, n_expert}, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {moe_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, moe_n_embd, n_expert}, 0); + layer.ffn_latent_up = create_tensor(tn(LLM_TENSOR_FFN_LATENT_UP, "weight", i), {moe_n_embd, n_embd}, 0); + } else { + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0); + } + } +} + +std::unique_ptr llama_model_maion_coder::build_arch_graph(const llm_graph_params & params) const { + return std::make_unique(*this, params); +} + +llama_model_maion_coder::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_k(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_v()); + GGML_ASSERT(n_embd_head == n_rot); + + ggml_tensor * inpL = build_inp_embd(model.tok_embd); + ggml_tensor * inp_pos = build_inp_pos(); + auto * inp_attn = build_attn_inp_kv_iswa(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + const float kq_scale = 1.0f / sqrtf(float(n_embd_head)); + + for (int il = 0; il < n_layer; ++il) { + // export per-layer input hidden state so speculative drafts (e.g. dflash/eagle3) can read it + res->t_layer_inp[il] = inpL; + + ggml_tensor * inpSA = inpL; + ggml_tensor * cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur, n_embd_head, n_head, n_head_kv, il); + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + + // Full-attention layers do not use RoPE. + if (hparams.is_swa(il)) { + const float freq_base_l = model.get_rope_freq_base(cparams, il); + const float freq_scale_l = model.get_rope_freq_scale(cparams, il); + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, 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_l, freq_scale_l, ext_factor, attn_factor, beta_fast, beta_slow); + } + + cur = build_attn(inp_attn, model.layers[il].wo, NULL, model.layers[il].wo_s, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, 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); + } + + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + + if (model.layers[il].ffn_gate_inp == nullptr) { + 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); + } else { + ggml_tensor * router_logits = build_lora_mm(model.layers[il].ffn_gate_inp, cur); + ggml_tensor * inp_latent = build_lora_mm(model.layers[il].ffn_latent_down, cur); + cur = build_moe_ffn(inp_latent, + 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, hparams.n_expert_used(il), + LLM_FFN_SILU, true, 0.0f, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il, router_logits); + cur = build_lora_mm(model.layers[il].ffn_latent_up, cur); + } + + cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il); + cur = ggml_add(ctx0, cur, ffn_inp); + cur = build_cvec(cur, il); + inpL = cur; + } + + ggml_tensor * cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1); + cb(cur, "result_norm", -1); + res->t_embd = cur; + cur = build_lora_mm(model.output, cur, model.output_s); + cb(cur, "result_output", -1); + res->t_logits = cur; +} diff --git a/src/models/models.h b/src/models/models.h index a55983cfb..88fd07d97 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -196,6 +196,19 @@ struct llama_model_maincoder : public llama_model_base { }; +struct llama_model_maion_coder : public llama_model_base { + llama_model_maion_coder(const struct llama_model_params & params) : llama_model_base(params) {} + void load_arch_hparams(llama_model_loader & ml) override; + void load_arch_tensors(llama_model_loader & ml) override; + + struct graph : public llm_graph_context { + graph(const llama_model & model, const llm_graph_params & params); + }; + + std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; +}; + + struct llama_model_talkie : public llama_model_base { llama_model_talkie(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; diff --git a/tests/test-llama-archs.cpp b/tests/test-llama-archs.cpp index 1e8e915b1..3dd8f0655 100644 --- a/tests/test-llama-archs.cpp +++ b/tests/test-llama-archs.cpp @@ -135,6 +135,9 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) { uint32_t n_layer = 2; if (arch == LLM_ARCH_LLAMA4) { n_layer = 4; // hparams.n_no_rope_layer_step is hard-coded to 4 + } else if (arch == LLM_ARCH_MAION_CODER) { + n_head = 4; + n_layer = 4; // dense and MoE layers, each with local and full attention } else if (arch == LLM_ARCH_GEMMA4) { n_embd = 128; n_head = 2; @@ -182,7 +185,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) { uint32_t n_head_kv = n_head; if (arch == LLM_ARCH_QWEN3) { n_head_kv = 1; // MQA coverage - } else if (arch == LLM_ARCH_MUSE_GLIMMER || arch == LLM_ARCH_AFMOE) { + } else if (arch == LLM_ARCH_MUSE_GLIMMER || arch == LLM_ARCH_AFMOE || arch == LLM_ARCH_MAION_CODER) { n_head_kv = 2; // GQA coverage } const uint32_t n_embd_head = n_embd / n_head; @@ -289,7 +292,9 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) { ms.add_kv(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, uint32_t(8)); ms.add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW, n_ctx/8); - if (arch == LLM_ARCH_GEMMA4) { + if (arch == LLM_ARCH_MAION_CODER) { + ms.add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, std::vector{1, 1, 0, 0}); + } else if (arch == LLM_ARCH_GEMMA4) { ms.add_kv(LLM_KV_EMBEDDING_LENGTH_PER_LAYER, n_embd/2); ms.add_kv(LLM_KV_ATTENTION_SHARED_KV_LAYERS, uint32_t(0)); ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH_SWA, n_embd_head); @@ -443,6 +448,12 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) { } } + if (arch == LLM_ARCH_MAION_CODER) { + ms.add_kv(LLM_KV_MOE_LATENT_SIZE, n_embd / 2); + ms.add_kv(LLM_KV_EXPERT_COUNT, uint32_t(4)); + ms.add_kv(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, std::vector{n_ff, n_ff / 2, n_ff, n_ff}); + } + ms.add_kv(LLM_KV_POSNET_EMBEDDING_LENGTH, n_embd); ms.add_kv(LLM_KV_POSNET_BLOCK_COUNT, n_layer); ms.add_kv(LLM_KV_CONVNEXT_EMBEDDING_LENGTH, n_embd); @@ -654,6 +665,7 @@ static bool check_causal_attn_toggle( static bool moe_mandatory(const llm_arch arch) { switch (arch) { case LLM_ARCH_LLAMA4: + case LLM_ARCH_MAION_CODER: case LLM_ARCH_COHERE2MOE: case LLM_ARCH_GROK: case LLM_ARCH_QWEN2MOE: