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 10, 02:45 UTC
[ROCm][Bugfix] Disable hip sampler to fix deepseek's accuracy issue on ROCm (#32413)
Signed-off-by: ganyi <ygan@amd.com>
Pleaplusone committed
Jan 16, 2026 at 00:35 UTC
77c16df31df16d09b2198b485a6d54cc14e98b69
1 file changed
+4
vllm/v1/sample/ops/topk_topp_sampler.py
+4
@@ -174,6 +174,8 @@ class TopKTopPSampler(nn.Module):
174
k: torch.Tensor | None,
175
p: torch.Tensor | None,
176
) -> tuple[torch.Tensor, torch.Tensor | None]:
177
+ # FIXME: Fix aiter_sampler's accuracy issue and remove this flag
178
+ DISABLE_AITER_SAMPLER = True
179
"""Optimized ROCm/aiter path (same structure as forward_cuda)."""
180
if (k is None and p is None) or generators:
181
if generators:
@@ -186,6 +188,8 @@ class TopKTopPSampler(nn.Module):
188
"processed_logits",
189
"processed_logprobs",
190
), "aiter sampler does not support returning logits/logprobs."
191
+ if DISABLE_AITER_SAMPLER:
192
+ return self.forward_native(logits, generators, k, p)
193
return self.aiter_sample(logits, k, p, generators), None
194
195
def aiter_sample(