[Bugfix] Re-enable use_cudagraph in vLLM v1 (#19299)
Signed-off-by: Richard Zou <zou3519@gmail.com>
This commit is contained in:
@@ -15,7 +15,7 @@ class CompilationCounter:
|
||||
# not including the splitting ops
|
||||
num_piecewise_capturable_graphs_seen: int = 0
|
||||
num_backend_compilations: int = 0
|
||||
num_cudagraph_caputured: int = 0
|
||||
num_cudagraph_captured: int = 0
|
||||
# InductorAdapter.compile calls
|
||||
num_inductor_compiles: int = 0
|
||||
# EagerAdapter.compile calls
|
||||
|
||||
@@ -193,7 +193,7 @@ class CUDAPiecewiseBackend:
|
||||
entry.output = weak_ref_tensors(output)
|
||||
entry.cudagraph = cudagraph
|
||||
|
||||
compilation_counter.num_cudagraph_caputured += 1
|
||||
compilation_counter.num_cudagraph_captured += 1
|
||||
|
||||
# important: we need to return the output, rather than
|
||||
# the weak ref of the output, so that pytorch can correctly
|
||||
|
||||
Reference in New Issue
Block a user