A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 8, 14:13 UTC
[Bugfix] Fix FI kernel`chunk_gated_delta_rule` output shape for Qwen3.5 (#34219)
Signed-off-by: Roger Wang <hey@rogerw.io>
Roger Wang committed
Feb 10, 2026 at 02:41 UTC
ae4e280602f3c91d322a449f33f5aebbdd59ccc1
1 file changed
+3
-1
vllm/model_executor/models/qwen3_next.py
+3
-1
@@ -135,7 +135,7 @@ def fi_chunk_gated_delta_rule(
135
fi_state = initial_state.to(torch.float32)
136
fi_g = g.to(torch.float32)
137
fi_beta = beta.to(torch.float32)
138
- return chunk_gated_delta_rule_fi(
138
+ output, final_state = chunk_gated_delta_rule_fi(
139
q=q,
140
k=k,
141
v=v,
@@ -145,6 +145,8 @@ def fi_chunk_gated_delta_rule(
145
output_final_state=output_final_state,
146
cu_seqlens=cu_seqlens,
147
)
148
+ # Unsqueeze back to 4D (1, L, H, D) to match fla output format
149
+ return output.unsqueeze(0), final_state
150
151
152
@CustomOp.register("chunk_gated_delta_rule")