[GH-ISSUE #163] H3 (MiniMax) on RTX 5060 Ti (sm_120): text conditioning ignored + numerical blowup in DiT blocks #50

New issue

Open

opened 2026-10-26 22:19:42 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:19:42 +00:00

Owner

Originally created by @teemaker-dev on GitHub (Sep 7, 2026).
Original GitHub issue: https://github.com/Comfy-Org/comfy-kitchen/issues/163

H3 (MiniMax) on RTX 5060 Ti (sm_120): text conditioning ignored + numerical blowup in DiT blocks

Environment

GPU: RTX 5060 Ti 16GB (sm_120 Blackwell)
OS: Windows 11 + WSL2
ComfyUI 0.34.2 (Windows native)
torch 2.12.0+cu130 (also reproduced on 2.9.1+cu130)
comfy-kitchen 0.2.31

Models

minimax_h3_fl2va_pruned_w4a8_mixed.safetensors (12.5GB, W4A8)
minimax_h3_fl2va_pruned_int8_convrot.safetensors (20.9GB, INT8)

Setup

krea2 CLIP (qwen3vl_4b_fp8_scaled) + mmh3-4b-ClipProj-celeb-mlp projection, BasicGuider + BasicScheduler(simple, 4 steps) + SamplerCustomAdvanced + MiniMaxH3SigmaShift(12, 3), turbo 4-step LoRA, 768x432x124 frames.

Symptoms

Output is a fixed woven texture pattern, not the prompted scene. Both "mountain lake" and "sports car" prompts yield byte-identical frames (MD5 match at frames 0/30/60/90).
Different prompts → identical output, even though the context tensors differ (ctxsum -5617 vs -3591). Conditioning injection path verified working.
W4A8 and INT8 checkpoints → identical output frames (same MD5), suggesting the model is not actually driven by its weights.
Numerical blowup: after the 84 DiT blocks, text rows reach std ≈ 32256 (input std ≈ 2.26), video rows std ≈ 6944. Weights load correctly (block0 qkv_proj std ≈ 0.084, condition_proj 5376x5120).

Debug trail (all verified with instrumentation)

On torch 2.9.1 the model degraded to an identity mapping (DiT block outputs ≈ 0, latent never updated — sampling "ran" but output was the empty-latent VAE decode). Upgrading to torch 2.12.0 fixed this: the model now produces real dynamic output, but still ignores text.
Text conditioning verified: layout segments [(0,20,'text'), (20,434,'audio'), (434,12866,'video')], text_states (20,5376) injected at h[0:20] correctly, context content differs per prompt.
Standalone kernel tests pass: rms_rope_split_half_ (eager backend), torch SDPA, int8_linear (rel diff 0.18% vs eager).
Zeroing the context produces byte-identical output → text has zero influence on generation.
SDXL generation works fine on the same machine (does not use comfy-kitchen quantized ops).

#120 (same GPU family, GPU LOST under comfy-kitchen)
#55 (INT8 corrupted outputs on SM75)

Any hints? Is this a known incompatibility between Blackwell (sm_120) and comfy-kitchen quantized ops inside the H3 DiT? Happy to run any additional experiments or provide logs.

Originally created by @teemaker-dev on GitHub (Sep 7, 2026). Original GitHub issue: https://github.com/Comfy-Org/comfy-kitchen/issues/163 # H3 (MiniMax) on RTX 5060 Ti (sm_120): text conditioning ignored + numerical blowup in DiT blocks ## Environment - GPU: RTX 5060 Ti 16GB (sm_120 Blackwell) - OS: Windows 11 + WSL2 - ComfyUI 0.34.2 (Windows native) - torch 2.12.0+cu130 (also reproduced on 2.9.1+cu130) - comfy-kitchen 0.2.31 ## Models - `minimax_h3_fl2va_pruned_w4a8_mixed.safetensors` (12.5GB, W4A8) - `minimax_h3_fl2va_pruned_int8_convrot.safetensors` (20.9GB, INT8) ## Setup krea2 CLIP (`qwen3vl_4b_fp8_scaled`) + `mmh3-4b-ClipProj-celeb-mlp` projection, BasicGuider + BasicScheduler(simple, 4 steps) + SamplerCustomAdvanced + MiniMaxH3SigmaShift(12, 3), turbo 4-step LoRA, 768x432x124 frames. ## Symptoms 1. Output is a fixed woven texture pattern, not the prompted scene. Both "mountain lake" and "sports car" prompts yield **byte-identical frames** (MD5 match at frames 0/30/60/90). 2. Different prompts → identical output, **even though the context tensors differ** (ctxsum -5617 vs -3591). Conditioning injection path verified working. 3. W4A8 and INT8 checkpoints → **identical output frames** (same MD5), suggesting the model is not actually driven by its weights. 4. **Numerical blowup**: after the 84 DiT blocks, text rows reach std ≈ 32256 (input std ≈ 2.26), video rows std ≈ 6944. Weights load correctly (block0 qkv_proj std ≈ 0.084, condition_proj 5376x5120). ## Debug trail (all verified with instrumentation) - On torch 2.9.1 the model degraded to an **identity mapping** (DiT block outputs ≈ 0, latent never updated — sampling "ran" but output was the empty-latent VAE decode). Upgrading to torch 2.12.0 fixed this: the model now produces real dynamic output, but still ignores text. - Text conditioning verified: layout segments `[(0,20,'text'), (20,434,'audio'), (434,12866,'video')]`, text_states (20,5376) injected at h[0:20] correctly, context content differs per prompt. - Standalone kernel tests pass: `rms_rope_split_half_` (eager backend), torch SDPA, int8_linear (rel diff 0.18% vs eager). - Zeroing the context produces byte-identical output → text has **zero** influence on generation. - SDXL generation works fine on the same machine (does not use comfy-kitchen quantized ops). ## Related - #120 (same GPU family, GPU LOST under comfy-kitchen) - #55 (INT8 corrupted outputs on SM75) Any hints? Is this a known incompatibility between Blackwell (sm_120) and comfy-kitchen quantized ops inside the H3 DiT? Happy to run any additional experiments or provide logs.

bsakaguchi referenced this issue

2026-10-26 22:20:25 +00:00

[PR #50] Try to optimize sdvquant implementation a bit. #105