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 4, 02:30 UTC
[CI][Examples][RLHF] Disable async scheduling in rlhf_async_new_apis (#42042)
Signed-off-by: haosdent <haosdent@gmail.com>
haosdent committed
May 8, 2026 at 19:58 UTC
52458b60a88310a7f730f067e5a431a55cd34afd
1 file changed
+6
examples/rl/rlhf_async_new_apis.py
+6
@@ -227,6 +227,9 @@ llm_kwargs = dict(
227
attention_backend=ATTN_BACKEND,
228
gpu_memory_utilization=0.75,
229
weight_transfer_config=WeightTransferConfig(backend="nccl"),
230
+ # TODO(haosdent): re-enable once #42043 is fixed. Both LLM
231
+ # instances must match.
232
+ async_scheduling=False,
233
)
234
llm_kwargs.update(rocm_determinism_kwargs)
235
@@ -364,6 +367,9 @@ llm_v2_kwargs = dict(
367
gpu_memory_utilization=0.75,
368
distributed_executor_backend="ray",
369
attention_backend=ATTN_BACKEND,
370
+ # TODO(haosdent): re-enable once #42043 is fixed. Both LLM
371
+ # instances must match.
372
+ async_scheduling=False,
373
)
374
llm_v2_kwargs.update(rocm_determinism_kwargs)
375