[Model][Dflash] Enable Dflash support for Qwen3NextForCausalLM targets (#45319)
Signed-off-by: Jonas I. Liechti <j-i-l@t4d.ch>
This commit is contained in:
@@ -1425,6 +1425,14 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
|
|||||||
max_model_len=8192, # Reduce max len to ensure test runs in low-VRAM CI env
|
max_model_len=8192, # Reduce max len to ensure test runs in low-VRAM CI env
|
||||||
max_num_seqs=32,
|
max_num_seqs=32,
|
||||||
),
|
),
|
||||||
|
"DFlashQwen3NextDraftModel": _HfExamplesInfo(
|
||||||
|
"Qwen/Qwen3-Coder-Next",
|
||||||
|
speculative_model="z-lab/Qwen3-Coder-Next-DFlash",
|
||||||
|
use_original_num_layers=True, # DFlash requires all layers
|
||||||
|
max_model_len=8192, # Reduce for CI
|
||||||
|
max_num_seqs=32,
|
||||||
|
min_transformers_version="4.56.3", # Required for Qwen3Next
|
||||||
|
),
|
||||||
# [Eagle]
|
# [Eagle]
|
||||||
"EagleCohereForCausalLM": _HfExamplesInfo(
|
"EagleCohereForCausalLM": _HfExamplesInfo(
|
||||||
"/host/engines/cohere-moe",
|
"/host/engines/cohere-moe",
|
||||||
|
|||||||
@@ -68,6 +68,7 @@ from .interfaces import (
|
|||||||
HasInnerState,
|
HasInnerState,
|
||||||
IsHybrid,
|
IsHybrid,
|
||||||
MixtureOfExperts,
|
MixtureOfExperts,
|
||||||
|
SupportsEagle3,
|
||||||
SupportsLoRA,
|
SupportsLoRA,
|
||||||
SupportsPP,
|
SupportsPP,
|
||||||
)
|
)
|
||||||
@@ -758,6 +759,7 @@ class Qwen3NextForCausalLM(
|
|||||||
SupportsPP,
|
SupportsPP,
|
||||||
QwenNextMixtureOfExperts,
|
QwenNextMixtureOfExperts,
|
||||||
IsHybrid,
|
IsHybrid,
|
||||||
|
SupportsEagle3,
|
||||||
):
|
):
|
||||||
packed_modules_mapping = {
|
packed_modules_mapping = {
|
||||||
"qkv_proj": [
|
"qkv_proj": [
|
||||||
|
|||||||
Reference in New Issue
Block a user