From 23841bf8d8f87925dd63790169fc3c8589db3e06 Mon Sep 17 00:00:00 2001 From: Aryan Kumar Date: Mon, 19 Oct 2026 12:43:49 -0700 Subject: [PATCH] [docs]: refer to the OmniRef checkpoint without an internal step number --- docs/training/minimax_h3_fp4_qad.md | 6 +++--- .../minimax_h3/omniref_nvfp4_qad.yaml | 2 +- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/docs/training/minimax_h3_fp4_qad.md b/docs/training/minimax_h3_fp4_qad.md index 0ec3f18c0..aeb93884a 100644 --- a/docs/training/minimax_h3_fp4_qad.md +++ b/docs/training/minimax_h3_fp4_qad.md @@ -30,9 +30,9 @@ The runnable configs are: | V2 bf16 teacher / starting weights | `FastVideo/FastVideo-FastH3-8-Step-V2` | | NVFP4 consumer baseline | `FastVideo/FastVideo-FastH3-8-Step-V2-NVFP4-Consumer` | -The OmniRef config starts from the **calibrated step-3500 OmniRef checkpoint -used as QAD step 0**, not from the released `-v2` result; point `init_from` at -that checkpoint. The V2 config starts from the bf16 V2 snapshot. +The OmniRef config starts from the **calibrated OmniRef checkpoint used as QAD +step 0**, not from the released `-v2` result; point `init_from` at that +checkpoint. The V2 config starts from the bf16 V2 snapshot. The published V2.1 checkpoint is the **linears-only variant** of this schedule: NVFP4 W4A4 linears with BF16 VSA attention, 300 steps. The shared diff --git a/examples/train/configs/knowledge_distillation/minimax_h3/omniref_nvfp4_qad.yaml b/examples/train/configs/knowledge_distillation/minimax_h3/omniref_nvfp4_qad.yaml index d026f8618..77cd44c2f 100644 --- a/examples/train/configs/knowledge_distillation/minimax_h3/omniref_nvfp4_qad.yaml +++ b/examples/train/configs/knowledge_distillation/minimax_h3/omniref_nvfp4_qad.yaml @@ -11,7 +11,7 @@ # (V2.1 Consumer) profile, calibration, and the checkpoint contract. # # Released result: FastVideo/FastH3-Omni-NVFP4-QAD-v2 (step 120 of 200). -# `init_from` is the calibrated step-3500 OmniRef checkpoint used as QAD step 0, +# `init_from` is the calibrated OmniRef checkpoint used as QAD step 0, # not that released result. The calibrated amax table is vendored under # examples/train/calibration/fasth3/ and pinned by sha256 below. models: