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

[Misc] Remove pad_for_cudagraphs from config (#30143)

Browse Source 
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Lucas Wilkinson
2026-01-20 13:05:48 -07:00
committed by GitHub
parent 86c69dc54c
commit 2261340806
9 changed files with 157 additions and 136 deletions
Download Patch File Download Diff File Expand all files Collapse all files

47
vllm/config/compilation.py
View File

@@ -581,15 +581,6 @@ class CompilationConfig:
local_cache_dir: str = field(default=None, init=False) # type: ignore
"""local cache dir for each rank"""
bs_to_padded_graph_size: list[int] = field(
default=None, # type: ignore
init=False,
)
"""optimization:
Intuitively, bs_to_padded_graph_size should be dict[int, int].
since we know all keys are in a range [0, max_cudagraph_capture_size],
we can optimize it to list[int] for better lookup performance."""
# keep track of enabled and disabled custom ops
enabled_custom_ops: Counter[str] = field(default_factory=Counter, init=False)
"""custom ops that are enabled"""
@@ -639,7 +630,6 @@ class CompilationConfig:
"debug_dump_path",
"cache_dir",
"local_cache_dir",
"bs_to_padded_graph_size",
"traced_files",
"compilation_time",
"static_forward_context",
@@ -661,7 +651,6 @@ class CompilationConfig:
"enabled_custom_ops": True,
"disabled_custom_ops": True,
"compilation_time": True,
"bs_to_padded_graph_size": True,
"traced_files": True,
"inductor_compile_config": {
"post_grad_custom_post_pass": True,
@@ -882,7 +871,6 @@ class CompilationConfig:
"""To complete the initialization after cudagraph related
configs are set. This includes:
- initialize compile_sizes
- pre-compute the mapping bs_to_padded_graph_size
"""
computed_compile_sizes = []
@@ -906,23 +894,6 @@ class CompilationConfig:
if self.cudagraph_capture_sizes:
assert self.cudagraph_capture_sizes[-1] == self.max_cudagraph_capture_size
# May get recomputed in the model runner if adjustment is needed for spec-decode
self.compute_bs_to_padded_graph_size()
# Validate that compile_sizes won't be changed by padding.
# Only validate when cudagraphs are actually being used.
if self.compile_sizes and self.cudagraph_mode != CUDAGraphMode.NONE:
for size in self.compile_sizes:
if size <= self.max_cudagraph_capture_size:
padded = self.bs_to_padded_graph_size[size]
if padded != size:
raise ValueError(
f"compile_sizes contains {size} which would be "
f"padded to {padded}. All compile_sizes must be "
"values that won't be changed by cudagraph padding. "
"Use values from cudagraph_capture_sizes."
)
def set_splitting_ops_for_v1(
self, all2all_backend: str, data_parallel_size: int = 1
):
@@ -1134,24 +1105,6 @@ class CompilationConfig:
self.max_cudagraph_capture_size = rounded_sizes[-1]
self.cudagraph_capture_sizes = rounded_sizes
# Recompute after adjusting the cudagraph sizes
self.compute_bs_to_padded_graph_size()
def compute_bs_to_padded_graph_size(self):
# pre-compute the mapping from batch size to padded graph size
self.bs_to_padded_graph_size = [
0 for i in range(self.max_cudagraph_capture_size + 1)
]
for end, start in zip(
self.cudagraph_capture_sizes + [self.max_cudagraph_capture_size + 1],
[0] + self.cudagraph_capture_sizes,
):
for bs in range(start, end):
if bs == start:
self.bs_to_padded_graph_size[bs] = start
else:
self.bs_to_padded_graph_size[bs] = end
def get_compile_ranges(self) -> list[Range]:
"""Get the compile ranges for the compilation config."""
if self.compile_ranges_split_points is None: