Merge remote-tracking branch 'upstream/main' into patches

This commit is contained in:
2026-10-05 08:52:31 +07:00
85 changed files with 3146 additions and 769 deletions
+11 -8
View File
@@ -19,10 +19,13 @@ from vllm.compilation.breakable_cudagraph import (
)
from vllm.compilation.counter import compilation_counter
from vllm.compilation.cuda_graph import CUDAGraphStat, CUDAGraphWrapper
from vllm.compilation.cudagraph_pool import (
capture_outside_cumem_pool,
capture_pool,
)
from vllm.compilation.wrapper import TorchCompileWithNoGuardsWrapper
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.config.compilation import CUDAGraphMode
from vllm.distributed.device_communicators.pynccl_allocator import set_graph_pool_id
from vllm.distributed.parallel_state import (
get_pp_group,
graph_capture,
@@ -481,15 +484,14 @@ class CudaGraphManager:
# Sync offloader's copy stream before capture.
# Ensure any pre-capture prefetches from offloader are complete.
get_offloader().sync_prev_onload()
if self.pool is not None:
set_graph_pool_id(self.pool)
else:
set_graph_pool_id(current_platform.graph_pool_handle())
if self._capture_mem_samples is not None:
torch.accelerator.synchronize()
free_before = torch.accelerator.get_memory_info()[0]
with torch.cuda.graph(
graph, self.pool, stream=self._capture_stream(desc)
with (
capture_pool(self.pool, self.vllm_config) as pool,
torch.cuda.graph(
graph, pool, stream=self._capture_stream(desc)
),
):
forward_fn(CUDAGraphMode.NONE)
# Join offloader's copy stream after forward to avoid
@@ -941,7 +943,8 @@ def profile_cudagraph_memory(runner: "GPUModelRunner") -> int:
mem_samples: list[int] = []
manager._capture_mem_samples = mem_samples
measured = int(runner.capture_model(profile_only=True))
with capture_outside_cumem_pool():
measured = int(runner.capture_model(profile_only=True))
# The measured delta covers PIECEWISE, encoder and speculator graphs
# plus the sampled FULL graphs; swap the sampled FULL cost for the