[Bugfix][ROCm] Fix startup OOM in AITER MLA FP8 prefill workspace sizing (#57923)

Signed-off-by: simondanielsson <simon.danielsson99@hotmail.com>
This commit is contained in:
Simon Danielsson
2026-09-23 11:44:12 -05:00
committed by GitHub
parent c4f6ce4874
commit b5ea0c7107
@@ -772,9 +772,18 @@ class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
device=device,
)
from vllm.platforms import current_platform
from vllm.v1.worker.workspace import current_workspace_manager
max_num_partial_tiles = reduce_partial_map_size
# AITER metadata sizing assumes all requests can carry max_prefill_qlen tokens,
# which is a loose worst case in the number of QO tiles. Rather, the sum of
# qlens is bounded by max_num_batched_tokens. Manually compute the number of
# qo tiles to avoid OOM at startup.
# TODO: AITER should give us this budget constrained value
qo_tile_cnt = (
cdiv(max_num_batched_tokens, _FP8_PREFILL_TILE_Q) + max_num_reqs - 1
)
max_num_partial_tiles = qo_tile_cnt + current_platform.num_compute_units()
reservations: list[tuple[tuple[int, ...], torch.dtype]] = [
(
(max_num_partial_tiles * _FP8_PREFILL_TILE_Q, num_head_k, v_head_dim),