sleepy/llama.cpp
1
0
Fork 0
Code Issues 9 Pull Requests Actions 103 Packages Projects Releases Wiki Activity

models : support qwen3.5 series (#19468)

Browse Source 
* support qwen3.5 series

* remove deepstack for now, and some code clean

* code clean

* add FULL_ATTENTION_INTERVAL metadata

* code clean

* reorder v heads for linear attention to avoid expensive interleaved repeat
This commit is contained in:
JJJYmmm
2026-02-11 00:00:26 +08:00
committed by GitHub
parent 9a96352729
commit fc0fe40049
17 changed files with 2096 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
gguf-py/gguf/constants.py
+64 -2
View File
@@ -142,6 +142,7 @@ class Keys:
EMBEDDING_SCALE = "{arch}.embedding_scale"
TOKEN_SHIFT_COUNT = "{arch}.token_shift_count"
INTERLEAVE_MOE_LAYER_STEP = "{arch}.interleave_moe_layer_step"
FULL_ATTENTION_INTERVAL = "{arch}.full_attention_interval"
ACTIVATION_SPARSITY_SCALE = "{arch}.activation_sparsity_scale"
ALTUP_ACTIVE_IDX = "{arch}.altup.active_idx"
ALTUP_NUM_INPUTS = "{arch}.altup.num_inputs"
@@ -384,6 +385,8 @@ class MODEL_ARCH(IntEnum):
QWEN3NEXT = auto()
QWEN3VL = auto()
QWEN3VLMOE = auto()
QWEN35 = auto()
QWEN35MOE = auto()
PHI2 = auto()
PHI3 = auto()
PHIMOE = auto()
@@ -557,13 +560,14 @@ class MODEL_TENSOR(IntEnum):
SSM_D = auto()
SSM_NORM = auto()
SSM_OUT = auto()
SSM_ALPHA = auto() # qwen3.5
SSM_BETA_ALPHA = auto() # qwen3next
SSM_CONV1D_Q = auto() # Kimi Linear
SSM_CONV1D_K = auto() # Kimi Linear
SSM_CONV1D_V = auto() # Kimi Linear
SSM_F_A = auto() # Kimi Linear
SSM_F_B = auto() # Kimi Linear
SSM_BETA = auto() # Kimi Linear
SSM_BETA = auto() # Kimi Linear qwen3.5
SSM_G_A = auto() # Kimi Linear
SSM_G_B = auto() # Kimi Linear
TIME_MIX_W0 = auto()
@@ -814,6 +818,8 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.QWEN3NEXT: "qwen3next",
MODEL_ARCH.QWEN3VL: "qwen3vl",
MODEL_ARCH.QWEN3VLMOE: "qwen3vlmoe",
MODEL_ARCH.QWEN35: "qwen35",
MODEL_ARCH.QWEN35MOE: "qwen35moe",
MODEL_ARCH.PHI2: "phi2",
MODEL_ARCH.PHI3: "phi3",
MODEL_ARCH.PHIMOE: "phimoe",
@@ -985,13 +991,14 @@ 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_ALPHA: "blk.{bid}.ssm_alpha", # qwen3.5
MODEL_TENSOR.SSM_BETA_ALPHA: "blk.{bid}.ssm_ba",
MODEL_TENSOR.SSM_CONV1D_Q: "blk.{bid}.ssm_conv1d_q", # Kimi Linear
MODEL_TENSOR.SSM_CONV1D_K: "blk.{bid}.ssm_conv1d_k", # Kimi Linear
MODEL_TENSOR.SSM_CONV1D_V: "blk.{bid}.ssm_conv1d_v", # Kimi Linear
MODEL_TENSOR.SSM_F_A: "blk.{bid}.ssm_f_a", # Kimi Linear
MODEL_TENSOR.SSM_F_B: "blk.{bid}.ssm_f_b", # Kimi Linear
MODEL_TENSOR.SSM_BETA: "blk.{bid}.ssm_beta", # Kimi Linear
MODEL_TENSOR.SSM_BETA: "blk.{bid}.ssm_beta", # Kimi Linear qwen3.5
MODEL_TENSOR.SSM_G_A: "blk.{bid}.ssm_g_a", # Kimi Linear
MODEL_TENSOR.SSM_G_B: "blk.{bid}.ssm_g_b", # Kimi Linear
MODEL_TENSOR.TIME_MIX_W0: "blk.{bid}.time_mix_w0",
@@ -1818,6 +1825,61 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.QWEN35: [
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.ATTN_QKV,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.SSM_A,
MODEL_TENSOR.SSM_CONV1D,
MODEL_TENSOR.SSM_DT,
MODEL_TENSOR.SSM_NORM,
MODEL_TENSOR.SSM_BETA,
MODEL_TENSOR.SSM_ALPHA,
MODEL_TENSOR.SSM_OUT
],
MODEL_ARCH.QWEN35MOE: [
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.ATTN_QKV,
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_BETA,
MODEL_TENSOR.SSM_ALPHA,
MODEL_TENSOR.SSM_OUT
],
MODEL_ARCH.PLAMO: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
gguf-py/gguf/gguf_writer.py
+3
View File
@@ -708,6 +708,9 @@ class GGUFWriter:
def add_leading_dense_block_count(self, length: int) -> None:
self.add_uint32(Keys.LLM.LEADING_DENSE_BLOCK_COUNT.format(arch=self.arch), length)
def add_full_attention_interval(self, interval: int) -> None:
self.add_uint32(Keys.LLM.FULL_ATTENTION_INTERVAL.format(arch=self.arch), interval)
def add_feed_forward_length(self, length: int | Sequence[int]) -> None:
if isinstance(length, int):
self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length)
gguf-py/gguf/tensor_mapping.py
+8 -1
View File
@@ -228,6 +228,7 @@ class TensorNameMap:
"transformer_encoder.{bid}.qkv", # neobert
"layers.{bid}.attn.Wqkv", # modern-bert
"model.layers.{bid}.self_attn.language_expert_query_key_value", # cogvlm
"model.layers.{bid}.linear_attn.in_proj_qkv", # qwen3.5
),
# Attention query
@@ -359,6 +360,7 @@ class TensorNameMap:
MODEL_TENSOR.ATTN_GATE: (
"model.layers.{bid}.self_attn.gate_proj", # afmoe
"model.layers.{bid}.linear_attn.in_proj_z", # qwen3.5
"model.layers.{bid}.self_attn.g_proj", # step3.5 head-wise attention gate
),
@@ -823,6 +825,10 @@ class TensorNameMap:
"model.layers.layers.{bid}.mixer.out_proj", # plamo2
),
MODEL_TENSOR.SSM_ALPHA: (
"model.layers.{bid}.linear_attn.in_proj_a", # qwen3.5
),
MODEL_TENSOR.SSM_BETA_ALPHA: (
"model.layers.{bid}.linear_attn.in_proj_ba", # qwen3next
),
@@ -844,7 +850,8 @@ class TensorNameMap:
"model.layers.{bid}.self_attn.f_b_proj",
),
MODEL_TENSOR.SSM_BETA: (
"model.layers.{bid}.self_attn.b_proj",
"model.layers.{bid}.linear_attn.in_proj_b", # qwen3.5
"model.layers.{bid}.self_attn.b_proj", # Kimi Linear
),
MODEL_TENSOR.SSM_G_A: (
"model.layers.{bid}.self_attn.g_a_proj",