Skip to content

model: add DeepSeek V4 architecture #22607

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
nisparks wants to merge 1 commit into from
Closed

model: add DeepSeek V4 architecture #22607

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
243 changes: 210 additions & 33 deletions convert_hf_to_gguf.py
View file
Open in desktop

Large diffs are not rendered by default.

87 changes: 87 additions & 0 deletions gguf-py/gguf/constants.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -442,6 +442,7 @@ class MODEL_ARCH(IntEnum):
DEEPSEEK = auto()
DEEPSEEK2 = auto()
DEEPSEEK2OCR = auto()
DEEPSEEK4 = auto()
CHATGLM = auto()
GLM4 = auto()
GLM4_MOE = auto()
Expand Down Expand Up @@ -708,6 +709,27 @@ class MODEL_TENSOR(IntEnum):
INDEXER_PROJ = auto()
INDEXER_ATTN_K = auto()
INDEXER_ATTN_Q_B = auto()
ATTN_KV_LATENT = auto()
ATTN_OUT_A = auto()
ATTN_OUT_B = auto()
ATTN_COMPRESS_APE = auto()
ATTN_COMPRESS_NORM = auto()
ATTN_COMPRESS_KV = auto()
ATTN_COMPRESS_GATE = auto()
INDEXER_COMPRESS_APE = auto()
INDEXER_COMPRESS_NORM = auto()
INDEXER_COMPRESS_KV = auto()
INDEXER_COMPRESS_GATE = auto()
HC_HEAD_BASE = auto()
HC_HEAD_FN = auto()
HC_HEAD_SCALE = auto()
HC_ATTN_BASE = auto()
HC_ATTN_FN = auto()
HC_ATTN_SCALE = auto()
HC_FFN_BASE = auto()
HC_FFN_FN = auto()
HC_FFN_SCALE = auto()
FFN_GATE_TID2EID = auto()
# vision
V_MMPROJ = auto()
V_MMPROJ_FC = auto()
Expand Down Expand Up @@ -928,6 +950,7 @@ class MODEL_TENSOR(IntEnum):
MODEL_ARCH.DEEPSEEK: "deepseek",
MODEL_ARCH.DEEPSEEK2: "deepseek2",
MODEL_ARCH.DEEPSEEK2OCR: "deepseek2-ocr",
MODEL_ARCH.DEEPSEEK4: "deepseek4",
MODEL_ARCH.CHATGLM: "chatglm",
MODEL_ARCH.GLM4: "glm4",
MODEL_ARCH.GLM4_MOE: "glm4moe",
Expand Down Expand Up @@ -1193,6 +1216,27 @@ class MODEL_TENSOR(IntEnum):
MODEL_TENSOR.INDEXER_PROJ: "blk.{bid}.indexer.proj",
MODEL_TENSOR.INDEXER_ATTN_K: "blk.{bid}.indexer.attn_k",
MODEL_TENSOR.INDEXER_ATTN_Q_B: "blk.{bid}.indexer.attn_q_b",
MODEL_TENSOR.ATTN_KV_LATENT: "blk.{bid}.attn_kv_latent",
MODEL_TENSOR.ATTN_OUT_A: "blk.{bid}.attn_output_a",
MODEL_TENSOR.ATTN_OUT_B: "blk.{bid}.attn_output_b",
MODEL_TENSOR.ATTN_COMPRESS_APE: "blk.{bid}.attn_compress_ape",
MODEL_TENSOR.ATTN_COMPRESS_NORM: "blk.{bid}.attn_compress_norm",
MODEL_TENSOR.ATTN_COMPRESS_KV: "blk.{bid}.attn_compress_kv",
MODEL_TENSOR.ATTN_COMPRESS_GATE: "blk.{bid}.attn_compress_gate",
MODEL_TENSOR.INDEXER_COMPRESS_APE: "blk.{bid}.indexer.compress_ape",
MODEL_TENSOR.INDEXER_COMPRESS_NORM: "blk.{bid}.indexer.compress_norm",
MODEL_TENSOR.INDEXER_COMPRESS_KV: "blk.{bid}.indexer.compress_kv",
MODEL_TENSOR.INDEXER_COMPRESS_GATE: "blk.{bid}.indexer.compress_gate",
MODEL_TENSOR.HC_HEAD_BASE: "hc_head_base",
MODEL_TENSOR.HC_HEAD_FN: "hc_head_fn",
MODEL_TENSOR.HC_HEAD_SCALE: "hc_head_scale",
MODEL_TENSOR.HC_ATTN_BASE: "blk.{bid}.hc_attn_base",
MODEL_TENSOR.HC_ATTN_FN: "blk.{bid}.hc_attn_fn",
MODEL_TENSOR.HC_ATTN_SCALE: "blk.{bid}.hc_attn_scale",
MODEL_TENSOR.HC_FFN_BASE: "blk.{bid}.hc_ffn_base",
MODEL_TENSOR.HC_FFN_FN: "blk.{bid}.hc_ffn_fn",
MODEL_TENSOR.HC_FFN_SCALE: "blk.{bid}.hc_ffn_scale",
MODEL_TENSOR.FFN_GATE_TID2EID: "blk.{bid}.ffn_gate_tid2eid",
# vision
MODEL_TENSOR.V_MMPROJ: "mm.{bid}",
MODEL_TENSOR.V_MMPROJ_FC: "mm.model.fc",
Expand Down Expand Up @@ -2816,6 +2860,49 @@ class MODEL_TENSOR(IntEnum):
MODEL_TENSOR.FFN_UP_SHEXP,
MODEL_TENSOR.FFN_EXP_PROBS_B,
],
MODEL_ARCH.DEEPSEEK4: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.HC_HEAD_BASE,
MODEL_TENSOR.HC_HEAD_FN,
MODEL_TENSOR.HC_HEAD_SCALE,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q_A,
MODEL_TENSOR.ATTN_Q_B,
MODEL_TENSOR.ATTN_KV_LATENT,
MODEL_TENSOR.ATTN_Q_A_NORM,
MODEL_TENSOR.ATTN_KV_A_NORM,
MODEL_TENSOR.ATTN_OUT_A,
MODEL_TENSOR.ATTN_OUT_B,
MODEL_TENSOR.ATTN_SINKS,
MODEL_TENSOR.ATTN_COMPRESS_APE,
MODEL_TENSOR.ATTN_COMPRESS_NORM,
MODEL_TENSOR.ATTN_COMPRESS_KV,
MODEL_TENSOR.ATTN_COMPRESS_GATE,
MODEL_TENSOR.INDEXER_PROJ,
MODEL_TENSOR.INDEXER_ATTN_Q_B,
MODEL_TENSOR.INDEXER_COMPRESS_APE,
MODEL_TENSOR.INDEXER_COMPRESS_NORM,
MODEL_TENSOR.INDEXER_COMPRESS_KV,
MODEL_TENSOR.INDEXER_COMPRESS_GATE,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_GATE_TID2EID,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
MODEL_TENSOR.FFN_GATE_SHEXP,
MODEL_TENSOR.FFN_DOWN_SHEXP,
MODEL_TENSOR.FFN_UP_SHEXP,
MODEL_TENSOR.FFN_EXP_PROBS_B,
MODEL_TENSOR.HC_ATTN_BASE,
MODEL_TENSOR.HC_ATTN_FN,
MODEL_TENSOR.HC_ATTN_SCALE,
MODEL_TENSOR.HC_FFN_BASE,
MODEL_TENSOR.HC_FFN_FN,
MODEL_TENSOR.HC_FFN_SCALE,
],
MODEL_ARCH.ERNIE4_5_MOE: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
Expand Down
104 changes: 103 additions & 1 deletion gguf-py/gguf/tensor_mapping.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,6 +36,7 @@ class TensorNameMap:
"encoder", # neobert
"model.transformer.wte", # llada
"embed_tokens", # qwen3-embedding
"embed", # deepseek-v4
),

# Token type embeddings
Expand Down Expand Up @@ -196,6 +197,7 @@ class TensorNameMap:
"layers.{bid}.input_layernorm", # qwen3-embedding
"model.layers.{bid}.attention_layernorm", # apertus
"model.layers.{bid}.pre_attention_layernorm", # kormo
"layers.{bid}.attn_norm", # deepseek-v4
),

# Attention norm 2
Expand Down Expand Up @@ -357,6 +359,7 @@ class TensorNameMap:
MODEL_TENSOR.ATTN_SINKS: (
"model.layers.{bid}.self_attn.sinks", # openai-moe
"model.layers.{bid}.self_attn.attention_sink_bias", # mimov2
"layers.{bid}.attn.attn_sink", # deepseek-v4
),

MODEL_TENSOR.ATTN_GATE: (
Expand Down Expand Up @@ -390,7 +393,8 @@ class TensorNameMap:
"layers.{bid}.post_attention_layernorm", # qwen3-embedding
"model.layers.{bid}.feedforward_layernorm", # apertus
"model.layers.{bid}.pre_mlp_layernorm", # kormo
"layers.{bid}.mlp_norm" # modern-bert
"layers.{bid}.mlp_norm", # modern-bert
"layers.{bid}.ffn_norm", # deepseek-v4
),

# Pre feed-forward norm
Expand Down Expand Up @@ -441,6 +445,7 @@ class TensorNameMap:
"backbone.layers.{bid}.mixer.gate", # nemotron-h-moe
"model.layers.{bid}.moe.gate", # step3.5
"model.layers.{bid}.router.proj", # gemma4
"layers.{bid}.ffn.gate", # deepseek-v4
),

MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
Expand All @@ -458,6 +463,7 @@ class TensorNameMap:
"model.layers.{bid}.mlp.e_score_correction", # exaone-moe
"model.layers.{bid}.block_sparse_moe.gate.e_score_correction", # kimi
"model.layers.{bid}.moe.router_bias", # step3.5 expert selection bias
"layers.{bid}.ffn.gate.bias", # deepseek-v4
),

# Feed-forward up
Expand Down Expand Up @@ -513,6 +519,7 @@ class TensorNameMap:
"encoder.layers.{bid}.mlp.experts.mlp.w1", # nomic-bert-moe
"model.layers.{bid}.block_sparse_moe.experts.up", # smallthinker
"model.layers.{bid}.moe.up_proj", # step3.5
"layers.{bid}.ffn.experts.w3", # deepseek-v4 (merged)
),

MODEL_TENSOR.FFN_UP_SHEXP: (
Expand All @@ -525,6 +532,7 @@ class TensorNameMap:
"backbone.layers.{bid}.mixer.shared_experts.up_proj", # nemotron-h-moe
"model.layers.{bid}.block_sparse_moe.shared_experts.up_proj", # kimi
"model.layers.{bid}.share_expert.up_proj", # step3.5
"layers.{bid}.ffn.shared_experts.w3", # deepseek-v4
),

MODEL_TENSOR.FFN_UP_CHEXP: (
Expand Down Expand Up @@ -565,6 +573,7 @@ class TensorNameMap:
"model.layers.{bid}.feed_forward.experts.gate_proj", # llama4
"model.layers.{bid}.block_sparse_moe.experts.gate", # smallthinker
"model.layers.{bid}.moe.gate_proj", # step3.5
"layers.{bid}.ffn.experts.w1", # deepseek-v4 (merged)
),

MODEL_TENSOR.FFN_GATE_SHEXP: (
Expand All @@ -575,6 +584,7 @@ class TensorNameMap:
"layers.{bid}.shared_experts.w1", # mistral-large
"model.layers.{bid}.block_sparse_moe.shared_experts.gate_proj", # kimi
"model.layers.{bid}.share_expert.gate_proj", # step3.5
"layers.{bid}.ffn.shared_experts.w1", # deepseek-v4
),

MODEL_TENSOR.FFN_GATE_CHEXP: (
Expand Down Expand Up @@ -644,6 +654,7 @@ class TensorNameMap:
"model.layers.{bid}.block_sparse_moe.experts.down", # smallthinker
"model.layers.{bid}.moe.down_proj", # step3.5
"model.layers.{bid}.experts.down_proj", # gemma4
"layers.{bid}.ffn.experts.w2", # deepseek-v4 (merged)
),

MODEL_TENSOR.FFN_DOWN_SHEXP: (
Expand All @@ -656,6 +667,7 @@ class TensorNameMap:
"backbone.layers.{bid}.mixer.shared_experts.down_proj", # nemotron-h-moe
"model.layers.{bid}.block_sparse_moe.shared_experts.down_proj", # kimi
"model.layers.{bid}.share_expert.down_proj", # step3.5
"layers.{bid}.ffn.shared_experts.w2", # deepseek-v4
),

MODEL_TENSOR.FFN_DOWN_CHEXP: (
Expand Down Expand Up @@ -1064,11 +1076,13 @@ class TensorNameMap:
MODEL_TENSOR.ATTN_Q_A: (
"model.layers.{bid}.self_attn.q_a_proj", # deepseek2
"layers.{bid}.attention.wq_a", # mistral-large
"layers.{bid}.attn.wq_a", # deepseek-v4
),

MODEL_TENSOR.ATTN_Q_B: (
"model.layers.{bid}.self_attn.q_b_proj", # deepseek2
"layers.{bid}.attention.wq_b", # mistral-large
"layers.{bid}.attn.wq_b", # deepseek-v4
),

MODEL_TENSOR.ATTN_KV_A_MQA: (
Expand All @@ -1093,11 +1107,97 @@ class TensorNameMap:
MODEL_TENSOR.ATTN_Q_A_NORM: (
"model.layers.{bid}.self_attn.q_a_layernorm", # deepseek2
"layers.{bid}.attention.q_a_norm", # mistral-large
"layers.{bid}.attn.q_norm", # deepseek-v4
),

MODEL_TENSOR.ATTN_KV_A_NORM: (
"model.layers.{bid}.self_attn.kv_a_layernorm", # deepseek2
"layers.{bid}.attention.kv_a_norm", # mistral-large
"layers.{bid}.attn.kv_norm", # deepseek-v4
),

MODEL_TENSOR.ATTN_KV_LATENT: (
"layers.{bid}.attn.wkv", # deepseek-v4
),

MODEL_TENSOR.ATTN_OUT_A: (
"layers.{bid}.attn.wo_a", # deepseek-v4
),

MODEL_TENSOR.ATTN_OUT_B: (
"layers.{bid}.attn.wo_b", # deepseek-v4
),

MODEL_TENSOR.ATTN_COMPRESS_APE: (
"layers.{bid}.attn.compressor.ape", # deepseek-v4
),

MODEL_TENSOR.ATTN_COMPRESS_NORM: (
"layers.{bid}.attn.compressor.norm", # deepseek-v4
),

MODEL_TENSOR.ATTN_COMPRESS_KV: (
"layers.{bid}.attn.compressor.wkv", # deepseek-v4
),

MODEL_TENSOR.ATTN_COMPRESS_GATE: (
"layers.{bid}.attn.compressor.wgate", # deepseek-v4
),

MODEL_TENSOR.INDEXER_COMPRESS_APE: (
"layers.{bid}.attn.indexer.compressor.ape", # deepseek-v4
),

MODEL_TENSOR.INDEXER_COMPRESS_NORM: (
"layers.{bid}.attn.indexer.compressor.norm", # deepseek-v4
),

MODEL_TENSOR.INDEXER_COMPRESS_KV: (
"layers.{bid}.attn.indexer.compressor.wkv", # deepseek-v4
),

MODEL_TENSOR.INDEXER_COMPRESS_GATE: (
"layers.{bid}.attn.indexer.compressor.wgate", # deepseek-v4
),

MODEL_TENSOR.HC_HEAD_BASE: (
"hc_head_base", # deepseek-v4
),

MODEL_TENSOR.HC_HEAD_FN: (
"hc_head_fn", # deepseek-v4
),

MODEL_TENSOR.HC_HEAD_SCALE: (
"hc_head_scale", # deepseek-v4
),

MODEL_TENSOR.HC_ATTN_BASE: (
"layers.{bid}.hc_attn_base", # deepseek-v4
),

MODEL_TENSOR.HC_ATTN_FN: (
"layers.{bid}.hc_attn_fn", # deepseek-v4
),

MODEL_TENSOR.HC_ATTN_SCALE: (
"layers.{bid}.hc_attn_scale", # deepseek-v4
),

MODEL_TENSOR.HC_FFN_BASE: (
"layers.{bid}.hc_ffn_base", # deepseek-v4
),

MODEL_TENSOR.HC_FFN_FN: (
"layers.{bid}.hc_ffn_fn", # deepseek-v4
),

MODEL_TENSOR.HC_FFN_SCALE: (
"layers.{bid}.hc_ffn_scale", # deepseek-v4
),

MODEL_TENSOR.FFN_GATE_TID2EID: (
"layers.{bid}.ffn.gate.tid2eid", # deepseek-v4
),

MODEL_TENSOR.ATTN_SUB_NORM: (
Expand Down Expand Up @@ -1244,6 +1344,7 @@ class TensorNameMap:

MODEL_TENSOR.INDEXER_PROJ: (
"model.layers.{bid}.self_attn.indexer.weights_proj", # DSA
"layers.{bid}.attn.indexer.weights_proj", # deepseek-v4
),

MODEL_TENSOR.INDEXER_ATTN_K: (
Expand All @@ -1252,6 +1353,7 @@ class TensorNameMap:

MODEL_TENSOR.INDEXER_ATTN_Q_B: (
"model.layers.{bid}.self_attn.indexer.wq_b", # DSA
"layers.{bid}.attn.indexer.wq_b", # deepseek-v4
),

############################################################################
Expand Down
1 change: 1 addition & 0 deletions src/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,7 @@ add_library(llama
llama-kv-cache.cpp
llama-kv-cache-iswa.cpp
llama-memory.cpp
llama-memory-deepseek4.cpp
llama-memory-hybrid.cpp
llama-memory-hybrid-iswa.cpp
llama-memory-recurrent.cpp
Expand Down
Loading