biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[BUGFIX] Fix hipErrorIllegalState in Qwen3-Omni during startup profiling allow inference Omni on ROCM (#33077)

Browse Source 
Signed-off-by: JartX <sagformas@epdcenter.es>
This commit is contained in:
JartX
2026-02-01 14:36:25 +01:00
committed by GitHub
parent b5f8c3092d
commit cd86fff38f
1 changed files with 31 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
vllm/model_executor/models/qwen3_omni_moe_thinker.py
View File

@@ -907,13 +907,37 @@ class Qwen3Omni_VisionTransformer(nn.Module):
hidden_states = hidden_states + pos_embeds
rotary_pos_emb_cos, rotary_pos_emb_sin = self.rot_pos_emb(grid_thw)
cu_seqlens = torch.repeat_interleave(
grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
).cumsum(
dim=0,
dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
)
cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
# RDNA3 (gfx11) specific bug workaround: torch.repeat_interleave triggers
# kernel crashes. We attempt the operation and catch the RuntimeError
# to switch to a vectorized cumsum + searchsorted approach.
try:
cu_seqlens = torch.repeat_interleave(
grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
).cumsum(
dim=0,
dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
)
cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
except RuntimeError:
logger.warning(
"torch.repeat_interleave not executable, "
"switching to vectorized searchsorted implementation."
)
repeat_counts = grid_thw[:, 0]
values = grid_thw[:, 1] * grid_thw[:, 2]
repeat_cumsum = repeat_counts.cumsum(0)
total_items = repeat_cumsum[-1].item()
indices = torch.searchsorted(
repeat_cumsum,
torch.arange(total_items, device=grid_thw.device),
right=True,
)
cu_seqlens = values[indices].cumsum(
dim=0,
dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
)
cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
hidden_states = hidden_states.unsqueeze(1)
rotary_pos_emb_cos = rotary_pos_emb_cos.to(hidden_states.device)