Repository navigation
[Bugfix][Hardware][Intel] MiniMax-H3: reuse the tiled-VAE gather buffer across requests on XPU - #6996
Closed
heyuanliu-intel wants to merge 1 commit into
Closed
heyuanliu-intel wants to merge 1 commit into
heyuanliu-intel wants to merge 1 commit into
Conversation
heyuanliu-intel
requested review from
NickCao,
wtomin and
yenuo26
as code owners
September 3, 2026 08:27
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
|
This PR appears to be related to model: MinimaxH3. Model owners: @david6666666 @heyuanliu-intel, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer. Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment. |
Omni ReviewBot triage noteResolved as of |
…er across requests on XPU ## Purpose **Problem:** the tiled video VAE gathers every rank's decoded tiles with a variable-shape all-gather whose output is allocated afresh each call. After the per-request `empty_cache()` the next gather lands at a new device address. On XPU the collective backend registers a device-side resource per new receive address and does not reclaim it, so device memory outside `torch.xpu.memory_allocated()` grows every request until the card is full. (Mechanism inferred from intervention experiments, not observed directly.) **Fix:** keep one landing buffer per (shape, dtype, device) on the adapter and gather into it with `all_gather_into_tensor`; tiles are cloned out because the buffer is reused. Installed only when the adapter's device is XPU, bound on the checkpoint instance; other backends keep the checkpoint's own method. The buffer cache is uncapped on purpose: the key is the tile geometry, and evicting would re-register a new address. ## Test Plan **Reproduce:** 4 XPU cards, TP4/SP1, model-level offload on, send three identical t2va requests (1344x768, 209 frames, 2 steps) and sample each rank's non-PyTorch device memory between requests: it climbs monotonically while `memory_allocated()` stays flat. ```bash pytest -q tests/diffusion/models/minimax_h3/test_minimax_h3_vae_gather_reuse.py ``` **vLLM Version:** as pinned by `requirements/`. **vLLM-Omni Commit:** rebased onto current `main`. ## Test Result CPU tests pass (XPU-only install, tile order for even and uneven shares, buffer reuse at same geometry, tiles survive the next gather). On the 4-card A/B, non-PyTorch peak growth from request 1 to 2 dropped from 6617.69 MiB to 4008.76 MiB per rank (-39.4%); worst-rank free memory at request 2 rose from 2.93 MiB to 1662.50 MiB. Growth is reduced, not eliminated. Signed-off-by: heyuanliu-intel <heyuan.liu@intel.com>
heyuanliu-intel
force-pushed
the
pr/h3-klvae-gather-buffer-reuse
branch
from
September 3, 2026 08:38
8817877 to
72fe1b6
Compare
hsliuustc0106
added
bug
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Purpose
Problem: the tiled video VAE gathers every rank's decoded tiles with a variable-shape all-gather whose output is allocated afresh each call. After the per-request
empty_cache()the next gather lands at a new device address. On XPU the collective backend registers a device-side resource per new receive address and does not reclaim it, so device memory outsidetorch.xpu.memory_allocated()grows every request until the card is full. (Mechanism inferred from intervention experiments, not observed directly.)Fix: keep one landing buffer per (shape, dtype, device) on the adapter and gather into it with
all_gather_into_tensor; tiles are cloned out because the buffer is reused. Installed only when the adapter's device is XPU, bound on the checkpoint instance; other backends keep the checkpoint's own method. The buffer cache is uncapped on purpose: the key is the tile geometry, and evicting would re-register a new address.Test Plan
Reproduce: 4 XPU cards, TP4/SP1, model-level offload on, send three identical t2va requests (1344x768, 209 frames, 2 steps) and sample each rank's non-PyTorch device memory between requests: it climbs monotonically while
memory_allocated()stays flat.vLLM Version: as pinned by
requirements/. vLLM-Omni Commit: rebased onto currentmain.Test Result
CPU tests pass (XPU-only install, tile order for even and uneven shares, buffer reuse at same geometry, tiles survive the next gather). On the 4-card A/B, non-PyTorch peak growth from request 1 to 2 dropped from 6617.69 MiB to 4008.76 MiB per rank (-39.4%); worst-rank free memory at request 2 rose from 2.93 MiB to 1662.50 MiB. Growth is reduced, not eliminated.