One-line fp16 fix doubles Qwen-Image 2.1 speed on RTX 20-series in ComfyUI
xrailgun · reddit · 2026-11-04
A Reddit user found that Qwen-Image 2.1, FIBO2, and MOSS-Image are declared bf16/fp32-only in ComfyUI, so Turing GPUs (no bf16 support) fall back to slow fp32. After checking activation ranges (65k max), they confirmed fp16 is safe for these three models.
The fix: in comfy/supportedmodels.py, change supportedinferencedtypes = [torch.bfloat16, torch.float32] to [torch.bfloat16, torch.float16, torch.float32] for those models — or use PR #17522.
Caveats:
- Only apply to these three models; others may not be fp16-safe (e.g. MiniMax H3 overflows to NaN)
- It changes compute dtype only (activations/attention/norms), not weights, so it works with any quant format (verified with int8 convrot checkpoints)
- Independent of other recent tensor-performance unlocks for 20/30/40-series GPUs and likely stacks multiplicatively
More from Infra
- RTX 5090 Street Prices Top $9,000 as Nvidia GPU Costs Climb — teleforce · 2026-11-04
- Qdrant beats Meilisearch 2x faster and 9x less memory on production data — qdrant_engine · 2026-11-04
- Antseed hires BD lead to build a peer-to-peer open market for AI inference with 700+ models — ravikantagrawal · 2026-11-04
- Texas set to overtake Virginia as top US data center state, with most projects still early-stage — asusarla · 2026-11-04
- How one AI agent cut serving costs 37% on 1M weekly messages, in 13 production changes — AccBalanced · 2026-11-04
- DeepSeek V4.1 Flash on 4x RTX 3090: 524k context at 30-40 tok/s decode — chota23 · 2026-11-04