fix: #17592 (review)

convert_hf_to_gguf.py

@@ -5121,7 +5121,7 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
 @ModelBase.register("KimiLinearModel", "KimiLinearForCausalLM")
 class KimiLinearModel(TextModel):
     """Kimi-Linear model with hybrid MLA+KDA architecture"""
-    model_arch = gguf.MODEL_ARCH.KIMI
+    model_arch = gguf.MODEL_ARCH.KIMI_LINEAR
 
     _experts: list[dict[str, Tensor]] | None = None
 

gguf-py/gguf/constants.py

@@ -445,7 +445,7 @@ class MODEL_ARCH(IntEnum):
     MINIMAXM2        = auto()
     RND1             = auto()
     PANGU_EMBED      = auto()
-    KIMI             = auto()  # Kimi-Linear (hybrid MLA+KDA)
+    KIMI_LINEAR      = auto()  # Kimi-Linear (hybrid MLA+KDA)
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -830,7 +830,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.COGVLM:           "cogvlm",
     MODEL_ARCH.RND1:             "rnd1",
     MODEL_ARCH.PANGU_EMBED:      "pangu-embedded",
-    MODEL_ARCH.KIMI:             "kimi",
+    MODEL_ARCH.KIMI_LINEAR:      "kimi-linear",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -3096,7 +3096,7 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
-    MODEL_ARCH.KIMI: [
+    MODEL_ARCH.KIMI_LINEAR: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
         MODEL_TENSOR.OUTPUT,

gguf-py/gguf/tensor_mapping.py

@@ -1570,15 +1570,33 @@ class TensorNameMap:
         ),
 
         # Kimi Linear KDA (using SSM_ prefix for consistency)
-        MODEL_TENSOR.SSM_CONV1D_Q: ("model.layers.{bid}.self_attn.q_conv1d",),
-        MODEL_TENSOR.SSM_CONV1D_K: ("model.layers.{bid}.self_attn.k_conv1d",),
-        MODEL_TENSOR.SSM_CONV1D_V: ("model.layers.{bid}.self_attn.v_conv1d",),
-        MODEL_TENSOR.SSM_F_A:      ("model.layers.{bid}.self_attn.f_a_proj",),
-        MODEL_TENSOR.SSM_F_B:      ("model.layers.{bid}.self_attn.f_b_proj",),
-        MODEL_TENSOR.SSM_BETA:     ("model.layers.{bid}.self_attn.b_proj",),
-        MODEL_TENSOR.SSM_A_LOG:    ("model.layers.{bid}.self_attn.A_log",),
-        MODEL_TENSOR.SSM_G_A:      ("model.layers.{bid}.self_attn.g_a_proj",),
-        MODEL_TENSOR.SSM_G_B:      ("model.layers.{bid}.self_attn.g_b_proj",),
+        MODEL_TENSOR.SSM_CONV1D_Q: (
+            "model.layers.{bid}.self_attn.q_conv1d",
+        ),
+        MODEL_TENSOR.SSM_CONV1D_K: (
+            "model.layers.{bid}.self_attn.k_conv1d",
+        ),
+        MODEL_TENSOR.SSM_CONV1D_V: (
+            "model.layers.{bid}.self_attn.v_conv1d",
+        ),
+        MODEL_TENSOR.SSM_F_A: (
+            "model.layers.{bid}.self_attn.f_a_proj",
+        ),
+        MODEL_TENSOR.SSM_F_B: (
+            "model.layers.{bid}.self_attn.f_b_proj",
+        ),
+        MODEL_TENSOR.SSM_BETA: (
+            "model.layers.{bid}.self_attn.b_proj",
+        ),
+        MODEL_TENSOR.SSM_A_LOG: (
+            "model.layers.{bid}.self_attn.A_log",
+        ),
+        MODEL_TENSOR.SSM_G_A: (
+            "model.layers.{bid}.self_attn.g_a_proj",
+        ),
+        MODEL_TENSOR.SSM_G_B: (
+            "model.layers.{bid}.self_attn.g_b_proj",
+        ),
         MODEL_TENSOR.SSM_DT_B: (
             "model.layers.{bid}.self_attn.dt_bias",
         ),

src/CMakeLists.txt

@@ -82,7 +82,7 @@ add_library(llama
             models/internlm2.cpp
             models/jais.cpp
             models/jamba.cpp
-            models/kimi.cpp
+            models/kimi-linear.cpp
             models/lfm2.cpp
             models/llada-moe.cpp
             models/llada.cpp

src/llama-arch.cpp

@@ -111,7 +111,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_COGVLM,           "cogvlm"           },
     { LLM_ARCH_RND1,             "rnd1"             },
     { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
-    { LLM_ARCH_KIMI,             "kimi"             },
+    { LLM_ARCH_KIMI_LINEAR,      "kimi-linear"      },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -2494,7 +2494,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
         },
     },
     {
-        LLM_ARCH_KIMI,
+        LLM_ARCH_KIMI_LINEAR,
         {
             { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
             { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
@@ -2833,7 +2833,7 @@ bool llm_arch_is_recurrent(const llm_arch & arch) {
         case LLM_ARCH_RWKV6QWEN2:
         case LLM_ARCH_RWKV7:
         case LLM_ARCH_ARWKV7:
-        case LLM_ARCH_KIMI:  // KDA layers use delta attention with recurrent state
+        case LLM_ARCH_KIMI_LINEAR:  // KDA layers use delta attention with recurrent state
             return true;
         default:
             return false;
@@ -2852,7 +2852,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
         case LLM_ARCH_QWEN3NEXT:
         // Kimi: Currently using recurrent-only mode since MLA doesn't use KV cache
         // TODO: Enable hybrid when MLA KV caching is implemented
-        // case LLM_ARCH_KIMI:
+        // case LLM_ARCH_KIMI_LINEAR:
             return true;
         default:
             return false;

src/llama-arch.h

@@ -115,7 +115,7 @@ enum llm_arch {
     LLM_ARCH_COGVLM,
     LLM_ARCH_RND1,
     LLM_ARCH_PANGU_EMBED,
-    LLM_ARCH_KIMI,
+    LLM_ARCH_KIMI_LINEAR,
     LLM_ARCH_UNKNOWN,
 };
 

src/llama-context.cpp

@@ -1387,7 +1387,7 @@ void llama_context::output_reorder() {
 //
 
 uint32_t llama_context::graph_max_nodes() const {
-    if (model.arch == LLM_ARCH_QWEN3NEXT || model.arch == LLM_ARCH_KIMI) {
+    if (model.arch == LLM_ARCH_QWEN3NEXT || model.arch == LLM_ARCH_KIMI_LINEAR) {
         return std::max<uint32_t>(8192u, 32u*model.n_tensors());
     }
     return std::max<uint32_t>(1024u, 8u*model.n_tensors());

src/llama-model.cpp

@@ -2247,7 +2247,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
-        case LLM_ARCH_KIMI:
+        case LLM_ARCH_KIMI_LINEAR:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,    hparams.n_embd_head_k_mla, false);
@@ -6406,7 +6406,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
                     }
                 } break;
-            case LLM_ARCH_KIMI:
+            case LLM_ARCH_KIMI_LINEAR:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -7712,9 +7712,9 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_qwen3next>(*this, params);
             } break;
-        case LLM_ARCH_KIMI:
+        case LLM_ARCH_KIMI_LINEAR:
             {
-                llm = std::make_unique<llm_build_kimi>(*this, params);
+                llm = std::make_unique<llm_build_kimi_linear>(*this, params);
             } break;
         default:
             GGML_ABORT("fatal error");
@@ -7871,7 +7871,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_ARCTIC:
         case LLM_ARCH_DEEPSEEK:
         case LLM_ARCH_DEEPSEEK2:
-        case LLM_ARCH_KIMI:
+        case LLM_ARCH_KIMI_LINEAR:
         case LLM_ARCH_PLM:
         case LLM_ARCH_CHATGLM:
         case LLM_ARCH_GLM4:

src/llama-quant.cpp

@@ -726,7 +726,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     // sanity checks for models that have attention layers
     // Skip this check for Kimi models which have hybrid KDA+MLA architecture
     // (only MLA layers have attn_kv_b weights, KDA layers don't)
-    if (qs.n_attention_wv != 0 && !is_clip_model && model.arch != LLM_ARCH_KIMI)
+    if (qs.n_attention_wv != 0 && !is_clip_model && model.arch != LLM_ARCH_KIMI_LINEAR)
     {
         const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
         // attention layers have a non-zero number of kv heads

src/models/kimi.cpp → src/models/kimi-linear.cpp

@@ -1,6 +1,6 @@
 #include "models.h"
 
-llm_build_kimi::llm_build_kimi(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params), model(model) {
+llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params), model(model) {
     ggml_tensor * cur;
     ggml_tensor * inpL;
 

src/models/models.h

@@ -283,8 +283,8 @@ struct llm_build_jamba : public llm_graph_context_mamba {
     llm_build_jamba(const llama_model & model, const llm_graph_params & params);
 };
 
-struct llm_build_kimi : public llm_graph_context_mamba {
-    llm_build_kimi(const llama_model & model, const llm_graph_params & params);
+struct llm_build_kimi_linear : public llm_graph_context_mamba {
+    llm_build_kimi_linear(const llama_model & model, const llm_graph_params & params);
 private:
     const llama_model & model;
 };