[ROCm] Remove stale SDPA and skinny GEMM workarounds (#50907)

Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Andreas Karatzas
2026-08-11 14:19:16 -04:00
committed by GitHub
co-authored by OpenAI Codex
parent 4f2f31b82e
commit 1ab2801dde
56 changed files with 500 additions and 558 deletions
+27 -2
View File
@@ -6420,6 +6420,21 @@ class GPUModelRunner(
logits,
dummy_metadata,
)
# All-greedy is a separate Triton specialization from the
# mixed-sampling path above. Compile and exercise it before serving.
all_greedy_metadata = replace(
dummy_metadata,
temperature=None,
all_greedy=True,
all_random=False,
)
self.rejection_sampler(
dummy_spec_decode_metadata,
draft_probs,
logits,
all_greedy_metadata,
)
torch.accelerator.synchronize()
return sampler_output
def _dummy_pooler_run_task(
@@ -6615,16 +6630,26 @@ class GPUModelRunner(
@staticmethod
@contextmanager
def _freeze_gc():
gc_was_enabled = gc.isenabled()
gc.collect()
should_freeze = not envs.VLLM_ENABLE_CUDAGRAPH_GC
if should_freeze:
gc.freeze()
# A Triton kernel finalized during stream capture unloads its
# module and invalidates the captured graph.
gc.disable()
try:
yield
finally:
if should_freeze:
gc.unfreeze()
gc.collect()
try:
gc.unfreeze()
gc.collect()
finally:
if gc_was_enabled:
gc.enable()
else:
gc.disable()
def shutdown(self) -> None:
"""Release GPU tensors (model weights, KV caches, workspace) so that