Simplify the heuristics for setting --mem-fraction-static (#7054)

This commit is contained in:
Lianmin Zheng
2025-06-10 19:01:39 -07:00
committed by GitHub
parent 0f218731e3
commit 6b12d6a8d5
3 changed files with 57 additions and 50 deletions

View File

@@ -153,7 +153,7 @@ def get_batch_sizes_to_capture(model_runner: ModelRunner):
gpu_mem = get_device_memory_capacity()
if gpu_mem is not None:
if gpu_mem > 90 * 1024: # H200
if gpu_mem > 90 * 1024: # H200, H20
capture_bs += list(range(160, 257, 8))
if gpu_mem > 160 * 1000: # B200, MI300
capture_bs += list(range(256, 513, 16))