[feat] MiniMax H3: block-FP8 text encoder on Hopper, its converter, and a 4xH100 FastH3 profile

[stayinalive181]

Purpose

FastVideo gates the serialized block-FP8 Qwen3-VL text encoder on Blackwell, but only the FlashInfer groupwise GEMM is Blackwell-specific. On Hopper the encoder therefore cannot stay resident beside an FSDP-sharded DiT on 80 GB and is offloaded on every request: 1.82 s per clip on 8xH200 versus 0.04 s resident (hlander-ai/minimax-h3). This PR lets sm_90 load the same FP8 checkpoint: the weights stay FP8-resident (35.5 GB instead of 66.7 GB) and each language-model linear dequantizes its block-scaled weight to BF16 right before a plain matmul, which is the arithmetic the FlashInfer path performs internally. FP8 buys residency here rather than speed. The route and the converter recipe come from hlander-ai/minimax-h3 (Apache-2.0).

With that in place a 4xH100 FastH3 profile becomes runnable: FSDP-sharded DiT, ThunderKittens sm_90a VSA kernel, resident FP8 encoder.

Changes

• fastvideo/models/encoders/minimax_h3_checkpoint_fp8.py: get_min_capability 100 -> 90; validate_runtime no longer requires the Triton quantizer or the FlashInfer GEMM on sm_90; apply dequantizes through _dequantize_block_fp8_weight on sm_90. The Blackwell path is unchanged.
• scripts/checkpoint_conversion/quantize_minimax_h3_text_encoder_fp8.py: writes the serialized checkpoint the loader already reads (E4M3 weight plus [N/128, K/128] weight_scale_inv, quantization_config in config.json, regenerated index). There was no producer for this format in-tree.
• examples/inference/basic/basic_fasth3.py: --vsa-kernel tk (ThunderKittens sm_90a, FASTVIDEO_VSA_TK=1), --text-encoder-weights, --offload-text-encoder / --offload-vae (defaults unchanged), and an sm_90 extension check in validate_profile_dependencies.
• examples/inference/basic/basic_fasth3_h100.yaml and a "Hopper (H100 80 GB)" README section with the measured points.
• Tests: validate_runtime accepts (9, 0) without the Blackwell dependencies; the dequantizer is bit-exact against the reference broadcast and the FP8 round trip stays within E4M3 precision; the tk profile's environment, config and dependency check.

Test Plan

ruff check fastvideo/models/encoders/minimax_h3_checkpoint_fp8.py \
  fastvideo/tests/encoders/test_minimax_h3_qwen3_vl_checkpoint_fp8.py \
  fastvideo/tests/inference/test_basic_fasth3_profile.py \
  examples/inference/basic/basic_fasth3.py \
  scripts/checkpoint_conversion/quantize_minimax_h3_text_encoder_fp8.py
yapf --diff / codespell --toml pyproject.toml on the same files
python -m pytest fastvideo/tests/encoders/test_minimax_h3_qwen3_vl_checkpoint_fp8.py -q
python -m pytest fastvideo/tests/inference/test_basic_fasth3_profile.py -q

Test Results

Test output

ruff:      All checks passed! (the only report is the pre-existing SIM102 at minimax_h3_checkpoint_fp8.py:381, untouched)
yapf:      touched hunks clean (the file is outside pre-commit and was never yapf-formatted upstream)
codespell: clean
pytest fastvideo/tests/encoders/test_minimax_h3_qwen3_vl_checkpoint_fp8.py: 13 passed in 0.57s
pytest fastvideo/tests/inference/test_basic_fasth3_profile.py:                15 passed in 0.07s
(CPU, macOS arm64, torch 2.14.0)

Measured on FastVideo b2db0c0 with this encoder path (warm, loading and compilation excluded, 345 frames = 14.375 s of playback):

GPUs	Resolution	Generation	Source
8x H100 80 GB	1344x768	13.506 s (13.503 / 13.506 / 13.554)	hlander-ai/minimax-h3, Triton VSA
4x H100 80 GB, NVLink	960x544	12.98-14.02 s over 8 consecutive clips	Windflow streaming worker: TK VSA, FSDP DiT, resident FP8 encoder

The 4xH100 row was measured through a downstream streaming worker rather than basic_fasth3.py; basic_fasth3.py numbers for that shape on a fresh 4xH100 box can be added to this PR on request.

Checklist

• ruff, yapf and codespell run with the repository configuration on the touched files (pre-commit itself is not installed on the authoring machine)
• I added or updated tests for my changes
• I updated documentation if needed (README section, example YAML, scripts/checkpoint_conversion/AGENTS.md)
• I considered GPU memory impact of my changes (sm_90 only: one dequantized BF16 weight is materialized per linear call, at most ~0.26 GB for the largest layer, and released after the matmul; the resident encoder drops from 66.7 GB to 35.5 GB)

For model/pipeline changes, also check:

• SSIM regression not run here; the FP8 encoder is a precision change relative to the stock BF16 encoder on every architecture, and this PR does not change the Blackwell numerics
• Support matrix unchanged (Hopper is a new runtime for an existing model)

[mergify]

Merge Protections

🔴 1 of 1 protections blocking · waiting on 👀 reviews and 🤖 CI

	Protection	Waiting on
🔴	PR merge requirements	👀 reviews and 🤖 CI

🔴 PR merge requirements

Waiting for

• #approved-reviews-by>=1
• check-success=fastcheck-passed
• check-success=full-suite-passed

This rule is failing.

• #approved-reviews-by>=1
• check-success=fastcheck-passed
• check-success=full-suite-passed
• check-success~=pre-commit
• title~=(?i)^\[(feat|feature|bugfix|fix|refactor|perf|ci|doc|docs|misc|chore|kernel|new.?model|skill|skills|infra)\]