Deepseek V3 support added (#176)

Co-authored-by: Stanisław Szymczyk <[email protected]>

Author: saood06

convert_hf_to_gguf.py

@@ -590,6 +590,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "855059429035d75a914d1eda9f10a876752e281a054a7a3d421ef0533e5b6249":
             # ref: https://huggingface.co/HuggingFaceTB/SmolLM-135M
             res = "smollm"
+        if chkhsh == "877081d19cf6996e2c4ff0e1236341e9b7bde288f5311a56a937f0afbbb3aeb5":
+            # ref: https://huggingface.co/deepseek-ai/DeepSeek-V3
+            res = "deepseek-v3"
 
         if res is None:
             logger.warning("\n")

convert_hf_to_gguf_update.py

@@ -94,6 +94,7 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "codeshell",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
     {"name": "tekken",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
     {"name": "smollm",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", },
+    {"name": "deepseek-v3",    "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-V3"},
 ]
 
 

gguf-py/gguf/constants.py

@@ -89,6 +89,8 @@ class LLM:
         EXPERT_USED_COUNT                 = "{arch}.expert_used_count"
         EXPERT_SHARED_COUNT               = "{arch}.expert_shared_count"
         EXPERT_WEIGHTS_SCALE              = "{arch}.expert_weights_scale"
+        EXPERT_WEIGHTS_NORM               = "{arch}.expert_weights_norm"
+        EXPERT_GATING_FUNC                = "{arch}.expert_gating_func"
         POOLING_TYPE                      = "{arch}.pooling_type"
         LOGIT_SCALE                       = "{arch}.logit_scale"
         DECODER_START_TOKEN_ID            = "{arch}.decoder_start_token_id"
@@ -257,6 +259,7 @@ class MODEL_TENSOR(IntEnum):
     FFN_GATE_SHEXP       = auto()
     FFN_DOWN_SHEXP       = auto()
     FFN_UP_SHEXP         = auto()
+    FFN_EXP_PROBS_B      = auto()
     ATTN_Q_NORM          = auto()
     ATTN_K_NORM          = auto()
     LAYER_OUT_NORM       = auto()
@@ -387,6 +390,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_TENSOR.FFN_GATE_EXP:         "blk.{bid}.ffn_gate_exps",
     MODEL_TENSOR.FFN_DOWN_EXP:         "blk.{bid}.ffn_down_exps",
     MODEL_TENSOR.FFN_UP_EXP:           "blk.{bid}.ffn_up_exps",
+    MODEL_TENSOR.FFN_EXP_PROBS_B:      "blk.{bid}.exp_probs_b",
     MODEL_TENSOR.LAYER_OUT_NORM:       "blk.{bid}.layer_output_norm",
     MODEL_TENSOR.SSM_IN:               "blk.{bid}.ssm_in",
     MODEL_TENSOR.SSM_CONV1D:           "blk.{bid}.ssm_conv1d",
@@ -978,6 +982,7 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.FFN_GATE_SHEXP,
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B
     ],
     MODEL_ARCH.CHATGLM : [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -1177,6 +1182,10 @@ class GGMLQuantizationType(IntEnum):
     IQ2_TN  = 42,
 
 
+class ExpertGatingFuncType(IntEnum):
+    SOFTMAX  = 1
+    SIGMOID  = 2
+
 
 # TODO: add GGMLFileType from ggml_ftype in ggml.h
 

gguf-py/gguf/gguf_writer.py

@@ -26,6 +26,7 @@
     RopeScalingType,
     PoolingType,
     TokenType,
+    ExpertGatingFuncType,
 )
 
 from .quants import quant_shape_from_byte_shape
@@ -670,6 +671,12 @@ def add_expert_shared_count(self, count: int) -> None:
     def add_expert_weights_scale(self, value: float) -> None:
         self.add_float32(Keys.LLM.EXPERT_WEIGHTS_SCALE.format(arch=self.arch), value)
 
+    def add_expert_weights_norm(self, value: bool) -> None:
+        self.add_bool(Keys.LLM.EXPERT_WEIGHTS_NORM.format(arch=self.arch), value)
+
+    def add_expert_gating_func(self, value: ExpertGatingFuncType) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_GATING_FUNC.format(arch=self.arch), value.value)
+
     def add_layer_norm_eps(self, value: float) -> None:
         self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value)
 

gguf-py/gguf/tensor_mapping.py

@@ -251,6 +251,10 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.shared_expert_gate", # qwen2moe
         ),
 
+        MODEL_TENSOR.FFN_EXP_PROBS_B: (
+            "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3
+        ),
+
         # Feed-forward up
         MODEL_TENSOR.FFN_UP: (
             "gpt_neox.layers.{bid}.mlp.dense_h_to_4h",                # gptneox

include/llama.h

@@ -93,6 +93,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_TEKKEN         = 20,
         LLAMA_VOCAB_PRE_TYPE_SMOLLM         = 21,
         LLAMA_VOCAB_PRE_TYPE_CODESHELL      = 22,
+        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM  = 23, //llama.cpp lists this as 28
     };
 
     // note: these values should be synchronized with ggml_rope

src/llama-vocab.cpp

@@ -367,6 +367,13 @@ struct llm_tokenizer_bpe {
                     "\\p{N}+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM:
+                regex_exprs = {
+                    "\\p{N}{1,3}",
+                    "[一-龥぀-ゟ゠-ヿ]+",
+                    "[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\r\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\r\n]*|\\s*[\r\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER:
                 regex_exprs = {
                     "[\r\n]",