[Refactor] Deprecate head_first for chunk_gated_delta_rule (#34263)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2026-02-19 13:23:49 -05:00
committed by GitHub
parent 3eff45d793
commit c683d11c94
3 changed files with 8 additions and 36 deletions
@@ -867,7 +867,6 @@ class LlavaOnevisionForConditionalGeneration(nn.Module, SupportsMultiModal, Supp
mm_input_by_modality = self._parse_and_validate_multimodal_inputs(**kwargs)
if not mm_input_by_modality:
return []
return None
# The result multimodal_embeddings is tuple of tensors, with each
# tensor corresponding to a multimodal data item (image or video).
-6
View File
@@ -115,7 +115,6 @@ def fi_chunk_gated_delta_rule(
initial_state: torch.Tensor,
output_final_state: bool,
cu_seqlens: torch.LongTensor | None = None,
head_first: bool = False,
use_qk_l2norm_in_kernel: bool = True,
):
from flashinfer.gdn_prefill import (
@@ -172,7 +171,6 @@ class ChunkGatedDeltaRule(CustomOp):
initial_state: torch.Tensor,
output_final_state: bool,
cu_seqlens: torch.LongTensor | None = None,
head_first: bool = False,
use_qk_l2norm_in_kernel: bool = True,
):
return fi_chunk_gated_delta_rule(
@@ -184,7 +182,6 @@ class ChunkGatedDeltaRule(CustomOp):
initial_state=initial_state,
output_final_state=output_final_state,
cu_seqlens=cu_seqlens,
head_first=head_first,
use_qk_l2norm_in_kernel=use_qk_l2norm_in_kernel,
)
@@ -198,7 +195,6 @@ class ChunkGatedDeltaRule(CustomOp):
initial_state: torch.Tensor,
output_final_state: bool,
cu_seqlens: torch.LongTensor | None = None,
head_first: bool = False,
use_qk_l2norm_in_kernel: bool = True,
):
return fla_chunk_gated_delta_rule(
@@ -210,7 +206,6 @@ class ChunkGatedDeltaRule(CustomOp):
initial_state=initial_state,
output_final_state=output_final_state,
cu_seqlens=cu_seqlens,
head_first=head_first,
use_qk_l2norm_in_kernel=use_qk_l2norm_in_kernel,
)
@@ -790,7 +785,6 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
initial_state=initial_state,
output_final_state=True,
cu_seqlens=non_spec_query_start_loc,
head_first=False,
use_qk_l2norm_in_kernel=True,
)
# Init cache