From 98be07412ebcbb67be1aaaeb23c564cfa5931c02 Mon Sep 17 00:00:00 2001 From: Aryan Kumar Date: Sun, 1 Nov 2026 03:58:50 -0800 Subject: [PATCH] [misc]: regenerate env docs for NVFP4 FA4 Q/K transform and BF16 block variables --- docs/contributing/env_vars.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/docs/contributing/env_vars.md b/docs/contributing/env_vars.md index cb31fb74b..deed9ecbc 100644 --- a/docs/contributing/env_vars.md +++ b/docs/contributing/env_vars.md @@ -198,6 +198,8 @@ longer exists also fails the test, so the fixing pull request deletes its entry. | `FASTVIDEO_VSA_PTX_BLACKWELL` | bool | `0` | attention | VIDEO_SPARSE_ATTN_H3 sends tile-64 forwards to the data-center Blackwell (sm_100a / sm_103a) kernel; grad-tracking forwards need its autograd operator, else Triton. fastvideo-kernel reads the same variable with the same rule. Deprecated names: `FASTVIDEO_VSA_SM100A`. | | `FASTVIDEO_VSA_TRITON` | bool | `0` | attention | Force the Triton MiniMax-H3 sparse attention kernel. fastvideo-kernel reads the same variable. | | `FASTVIDEO_NVFP4_FA4` | bool | `0` | attention | FlashAttention-4 quantizes Q and K to NVFP4. An explicit nvfp4_fa4 attention implementation argument takes precedence. | +| `FASTVIDEO_NVFP4_FA4_QK_TRANSFORM` | one of none, smooth_k, hadamard, smooth_k+hadamard | `none` | attention | Exact Q/K preconditioning applied right before NVFP4 quantization on the FlashAttention-4 NVFP4 path: smooth_k subtracts each head's token-mean of K, hadamard rotates Q and K by the head-dim Hadamard matrix. Must match the qk_transform the model was QAD-trained with. An explicit nvfp4_qk_transform attention implementation argument takes precedence. | +| `FASTVIDEO_NVFP4_FA4_BF16_BLOCKS` | str | `""` | attention | Comma-separated MiniMax-H3 transformer_blocks indices whose attention stays in BF16 while FASTVIDEO_NVFP4_FA4 quantizes the rest, for example 45,49. | | `FASTVIDEO_FA4_PV_MODE` | one of bf16, fp8 | `bf16` | attention | V dtype on the NVFP4 FlashAttention-4 path (FLASH_ATTN with nvfp4_fa4, ATTN_QAT_INFER on sm_100a/sm_103a): bf16 keeps V in BF16; fp8 casts V to float8 e4m3 without scaling. An explicit fa4_pv_mode attention implementation argument takes precedence. | | `FASTVIDEO_DISABLE_ATTENTION_COMPILE` | bool | `1` | attention | Keep attention forward out of torch.compile graphs (torch.compiler.disable). Set it to 0 to let attention constructed under that setting be traced. Setting it explicitly to true also blocks regional compile. | | `FASTVIDEO_MLX_WINDOW` | int | `0` | attention | MLX FastWan windowed attention size in tokens. 0 uses full attention. |