[Hotfix] Fix required enable_mamba_track argument for Flashinfer autotune path (#15314)
This commit is contained in:
@@ -2305,6 +2305,7 @@ class ModelRunner:
|
||||
encoder_len_fill_value=0,
|
||||
num_tokens_per_bs=num_tokens_per_bs,
|
||||
cache_loc_dtype=torch.int64,
|
||||
enable_mamba_track=False,
|
||||
)
|
||||
buffers.num_token_non_padded[...] = num_tokens
|
||||
|
||||
|
||||
Reference in New Issue
Block a user