mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 34 Packages Projects Releases Wiki Activity

[Model] Qwen3.5 dense and MoE support (no vision) (#19435)

Browse Source 
* Unified delta net handling

* Remove old methods.

* Refactor and optimize

* Adapt autoregressive version from @ymcki

* Change to decay mask approach

* Fix bad permute

* Qwen 3.5 support

* Apply suggestions from code review

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Further fixes

* Use inheritance, remove unneeded conts

* Not like this!

* Remove ggml.h explicit import

* Remove transformers, fix the views

* ACTUALLY fix views, make super calls explicit in conversion.

* Fix conversion again

* Remove extra ggml.h imports

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This commit is contained in:
Piotr Wilkin (ilintar)
2026-02-09 00:24:08 +01:00
committed by GitHub
parent e06088da0f
commit 39bf692af1
14 changed files with 1532 additions and 399 deletions
Download Patch File Download Diff File Expand all files Collapse all files
gguf-py/gguf/constants.py
+59
View File
@@ -382,6 +382,8 @@ class MODEL_ARCH(IntEnum):
QWEN3 = auto()
QWEN3MOE = auto()
QWEN3NEXT = auto()
QWEN3_5 = auto()
QWEN3_5_MOE = auto()
QWEN3VL = auto()
QWEN3VLMOE = auto()
PHI2 = auto()
@@ -812,6 +814,8 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.QWEN3: "qwen3",
MODEL_ARCH.QWEN3MOE: "qwen3moe",
MODEL_ARCH.QWEN3NEXT: "qwen3next",
MODEL_ARCH.QWEN3_5: "qwen3_5",
MODEL_ARCH.QWEN3_5_MOE: "qwen3_5moe",
MODEL_ARCH.QWEN3VL: "qwen3vl",
MODEL_ARCH.QWEN3VLMOE: "qwen3vlmoe",
MODEL_ARCH.PHI2: "phi2",
@@ -1784,6 +1788,61 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.SSM_BETA_ALPHA,
MODEL_TENSOR.SSM_OUT
],
MODEL_ARCH.QWEN3_5: [
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_IN,
MODEL_TENSOR.SSM_BETA_ALPHA,
MODEL_TENSOR.SSM_OUT,
],
MODEL_ARCH.QWEN3_5_MOE: [
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_IN,
MODEL_TENSOR.SSM_BETA_ALPHA,
MODEL_TENSOR.SSM_OUT,
],
MODEL_ARCH.QWEN3VL: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
gguf-py/gguf/tensor_mapping.py
+4 -2
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
@@ -358,8 +359,9 @@ class TensorNameMap:
),
MODEL_TENSOR.ATTN_GATE: (
"model.layers.{bid}.self_attn.gate_proj", # afmoe
"model.layers.{bid}.self_attn.g_proj", # step3.5 head-wise attention gate
"model.layers.{bid}.self_attn.gate_proj", # afmoe
"model.layers.{bid}.self_attn.g_proj", # step3.5 head-wise attention gate
"model.layers.{bid}.linear_attn.in_proj_z", # qwen3.5
),
# Feed-forward norm