"refactor: refactor_repeated_interfaces" (#32486)

Signed-off-by: tom-zju <tanjianpingzju1990@gmail.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
This commit is contained in:
tjp_zju
2026-01-18 22:07:01 +08:00
committed by GitHub
co-authored by Wentao Ye
parent 38bf2ffb21
commit 2f03035a61
11 changed files with 43 additions and 77 deletions
@@ -25,6 +25,7 @@ from vllm.utils.deep_gemm import (
get_mk_alignment_for_contiguous_layout,
m_grouped_fp8_gemm_nt_contiguous,
)
from vllm.utils.math_utils import cdiv
def _generate_optimal_warmup_m_values(
@@ -40,9 +41,6 @@ def _generate_optimal_warmup_m_values(
device: The torch device to get properties from.
"""
def ceil_div(a: int, b: int) -> int:
return (a + b - 1) // b
# DeepGEMM's possible block sizes
block_ms = [64, 128, 256]
block_ns = list(range(16, min(257, n + 1), 16))
@@ -63,7 +61,7 @@ def _generate_optimal_warmup_m_values(
for wave in range(1, 11): # Up to 10 waves
# M where this block config transitions to next wave
target_blocks = wave * num_sms
m = target_blocks * block_m // ceil_div(n, block_n)
m = target_blocks * block_m // cdiv(n, block_n)
if 1 <= m <= max_tokens:
m_values.add(m)