nodes/hao-ai-lab-FastVideo
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[bugfix] set FASTVIDEO_H3_VAE_TILE_BATCH in tests through envs overrides

Browse Source 
The env-var policy test flagged the monkeypatch.setenv/delenv writes in the
H3 VAE tile-batch tests. Use envs.FASTVIDEO_H3_VAE_TILE_BATCH.override via
the env_overrides fixture, and regenerate the env table in
docs/contributing/env_vars.md for the new default and description.
This commit is contained in:
Aryan Kumar
2026-11-01 15:17:55 -08:00
parent 78a83c8276
commit 10e2e03c0f
3 changed files with 16 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/contributing/env_vars.md
+2 -2
View File
@@ -278,9 +278,9 @@ longer exists also fails the test, so the fixing pull request deletes its entry.
| `FASTVIDEO_H3_SPLICE_FROM_STEP` | int | `4` | eval | First denoising step run by FASTVIDEO_H3_SPLICE_TRANSFORMER. |
| `FASTVIDEO_H3_ENCODER_LAYERWISE` | bool | `0` | performance | Stream MiniMax-H3 text-encoder language layers through exact-size pinned host memory (text-only prompts). |
| `FASTVIDEO_H3_ENCODER_FUSED_DEQUANT` | bool | `0` | performance | Expand the serialized NVFP4 MiniMax-H3 text encoder with one fused Triton pass on GPUs without FP4 GEMM. |
| `FASTVIDEO_H3_VAE_TILE_BATCH` | int | `1` | performance | Spatial tiles per MiniMax-H3 video VAE decoder call; 1 decodes per tile. |
| `FASTVIDEO_H3_VAE_TILE_BATCH` | int | `8` | performance | Spatial tiles per MiniMax-H3 video VAE decoder call. The ViT decoder treats batch entries independently, so batching tiles matches per-tile decoding, except for NVFP4 decoders with dynamic activation scales, which always decode per tile. On GB200 at 768x1344, 8 tiles per call decodes 107 frames in 3.7 s instead of 5.4 s and 362 frames in 12.8 s instead of 18.3 s (+0.7 GiB peak). 1 decodes per tile. |
| `FASTVIDEO_H3_REF2VA_MEMO_ENTRIES` | int | `0` | performance | Entries in each content-keyed MiniMax-H3 memo of reference encodes (Qwen3-VL Ref2VA presentation, VAE keyframe latents); repeat references are reused exactly. 0 disables. |
| `FASTVIDEO_H3_VAE_TILE_PARALLEL` | bool | `0` | performance | Split MiniMax-H3 VAE spatial tiles (not only temporal chunks) across the sequence-parallel ranks in the parallel decode and keyframe encode; bitwise equal to the serial path. |
| `FASTVIDEO_H3_VAE_TILE_PARALLEL` | bool | `0` | performance | Split MiniMax-H3 VAE spatial tiles (not only temporal chunks) across the sequence-parallel ranks in the parallel decode and keyframe encode; bitwise equal to the serial path. Requires FASTVIDEO_H3_VAE_TILE_BATCH=1 (the split replays the per-tile path). |
| `FASTVIDEO_H3_VAE_NVFP4_FUSED` | bool | `1` | performance | Run NVFP4 MiniMax-H3 VAE decoder blocks through the bit-exact fused Triton kernels at inference. |
| `FASTVIDEO_H3_VAE_NVFP4_CUDA_GRAPH` | bool | `1` | performance | Replay small fused NVFP4 MiniMax-H3 VAE decoder calls (at most four 1797-token tiles) as per-shape CUDA graphs. |
| `FASTVIDEO_H3_VAE_NVFP4_DECODER` | path | unset | performance | Exported NVFP4 MiniMax-H3 video decoder (.pt, or .safetensors with bf16 masters or pre-packed NVFP4 weights) that replaces the VAE decoder at load; skips the INT8 ConvRot overlay. |