[Spec] Allow Qwen3.5 backends to opt into packed target verification (#43037)

This commit is contained in:
paulzhang-tm
2026-10-10 22:39:17 -07:00
committed by GitHub
parent 22d63f5e43
commit 5678b71b6c
+12 -6
View File
@@ -998,16 +998,23 @@ class Qwen3_5GatedDeltaNet(nn.Module):
)
use_strided_prefill_z = False
use_fused_decode_proj_conv = (
packed_verify = False
if forward_batch.forward_mode.is_target_verify():
from sglang.srt.model_executor.forward_context import get_attn_backend
backend = get_attn_backend()
backend = getattr(backend, "linear_attn_backend", backend)
packed_verify = getattr(backend, "supports_packed_verify", False)
use_fused_proj_conv = (
_gdn_decode_fused_proj_conv
and forward_batch.forward_mode.is_decode()
and (forward_batch.forward_mode.is_decode() or packed_verify)
and isinstance(projected_states_qkvz, torch.Tensor)
and isinstance(projected_states_ba, torch.Tensor)
)
use_fused_contiguous_unpack = (
self.num_v_heads // self.num_k_heads in _GDN_FUSED_QKVZBA_RATIOS
)
if use_fused_decode_proj_conv:
if use_fused_proj_conv:
# GDN owns indexed Conv1D state and the safe unpack/Conv boundary;
# it replaces these temporary B/A placeholders before recurrence.
mixed_qkv = (projected_states_qkvz, projected_states_ba)
@@ -1055,11 +1062,10 @@ class Qwen3_5GatedDeltaNet(nn.Module):
a=a,
b=b,
)
if use_fused_decode_proj_conv:
if use_fused_proj_conv:
if not isinstance(attn_result, tuple) or len(attn_result) != 2:
raise RuntimeError(
"Fused GDN decode projection/Conv1D backend must return "
"(core_attn_out, z)"
"Fused GDN projection/Conv1D backend must return (core_attn_out, z)"
)
core_attn_out, z = attn_result
else: