fix(OOM): skip flash_attn_varlen during profiling
Only patch_xformers_sdpa_seq.py changed. No yaml changes.
This commit is contained in:
@@ -200,6 +200,10 @@ FALLBACK_METHOD = '''
|
|||||||
max_seqlen = max(seq_lens_list)
|
max_seqlen = max(seq_lens_list)
|
||||||
|
|
||||||
try:
|
try:
|
||||||
|
# Skip flash_attn during profiling — OOMs on large dummy batch
|
||||||
|
import os
|
||||||
|
if os.environ.get("BI100_IN_STARTUP_PROFILE") == "1":
|
||||||
|
raise RuntimeError("skip flash_attn during profiling")
|
||||||
out = _ixf.flash_attn_varlen_func(
|
out = _ixf.flash_attn_varlen_func(
|
||||||
q_flat.to(torch.float16),
|
q_flat.to(torch.float16),
|
||||||
k_flat.to(torch.float16),
|
k_flat.to(torch.float16),
|
||||||
|
|||||||
Reference in New Issue
Block a user