[david6666666]

Summary

Fix intermittent MiniMax-H3 I2VA accuracy failures on H100/SM90 by matching the unfused BF16 dtype boundaries in the two fused RMSNorm + AdaLN modulation kernels on Hopper only.

Diagnosis

• The H100 I2VA failure is isolated to the repeated image-conditioned DiT trajectory; Ref2VA remains passing.
• Build 15562 with the separate VAE Q/K launches reverted from [Kernel][MiniMax-H3] Run Q/K RMSNorm-RoPE in one launch #7167 still failed, so the Q/K launch path is not the cause of this CI failure. That change has been reverted from this PR.
• The fused modulation kernels kept RMSNorm/AdaLN intermediates in FP32, while the unfused path rounds through BF16. On SM90, the repeated precision difference can move the 50-step I2VA result below the CI threshold.
• B300/SM103 keeps the existing FP32 path: enabling the BF16 path globally regressed the local B300 I2VA run, so this fix is intentionally architecture-scoped.

Fix

• Select the BF16 affine boundaries only when the CUDA compute capability major is 9.
• Keep the existing modulation path on other architectures.
• Register the shared modulation source in the MiniMax-H3 CI dependency map so this change runs the I2VA/Ref2VA accuracy job.

Validation

• B300, 4-GPU HSDP + 4-way Ulysses, static regional compile, existing path: SSIM 0.975250, PSNR 35.859771.
• The same B300 setup with the BF16 path forced globally: SSIM 0.969440, PSNR 33.222849, confirming the architecture gate is required.
• Targeted tests: 49 passed, 3 skipped.
• MiniMax-H3 I2VA/Ref2VA accuracy passed in Build 15553 and Build 15555; Build 15562 used the reverted Q/K launch experiment and failed.

Signed-off-by: david6666666 530634352@qq.com

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/diffusion/offloader.md, docs/design/module/diffusion/diffusion_model_integration.md, docs/design/module/diffusion/index.md.

Module owners: @wtomin @Isotr0py @princepride

Routing: @wtomin via module of the changed files, module named in the PR description, CODEOWNERS; @Isotr0py via module of the changed files, module named in the PR description; @princepride via module of the changed files, module named in the PR description

@david6666666, 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.

[vllm-omni-review-bot]

Omni ReviewBot triage note

Resolved as of 49fa48895100: the high-priority or low-quality signal noted on an earlier commit no longer applies.

[natureofnature]

The H100 result for 3b84cc4a is now available: Build 15549's accuracy job still fails I2VA with SSIM 0.971247 / PSNR 33.627597 dB (required PSNR ≥34), identical to the preceding failures.

One candidate worth checking is the keyframe cuDNN math mode. The current context still sets allow_tf32=True. Deterministic convolution does not imply that TF32 is disabled.

A controlled four-run comparison on 4×H800, using base cb439f3e8439 rather than this PR head, produced:

Run	Keyframe cuDNN TF32	SSIM	PSNR (dB)	Result
Original	True	0.947329	27.533157	Fail
Patched	False	0.973072	34.587187	Pass
Reverted	True	0.947329	27.533157	Fail
Patched again	False	0.973072	34.587187	Pass

The only model-code behavior change was allow_tf32=True → False inside _minimax_h3_keyframe_encode_context(); the parallel encoding path was unchanged. Generated MP4s were byte-identical within each setting, including after reverting. All four runs restarted the model service and completed the original I2VA fixture with unchanged inputs, seed 0, 50 steps, and thresholds. Offline asset loading preserved the fixture's original SHA256 checks.

Environment: MiniMax-H3 FL2VA revision 42ed227ee7df40d41602854ae760620d6eb651fe, PyTorch 2.13.0+cu130, cuDNN 92000, vLLM 0.29.0. Test:

python -m pytest -s -v -o addopts='' \
  tests/e2e/accuracy/minimax_h3/test_minimax_h3_i2va_ref2va_similarity.py::test_minimax_h3_i2va_matches_official_reference \
  --run-level full_model

This is evidence for one H800 fixture, not confirmation of an H100 fix or the official reference's math mode. Could we try the same scoped TF32 change on the failing H100 setup and compare the keyframe latents? The shared encode_image path also needs coverage for Ref2VA still-image inputs before generalizing the change.