Repository navigation
[Bug]: Calibration failure with Swift-1.5-Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp.gguf #11
Description
Version or commit
1.2
Environment
Windows 11
GF RTX 5070Ti
Steps to reproduce
strata calibrate --model Swift-1.5-Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp.gguf
Expected behavior
Should return the startup parameters to use, I assume.
Actual behavior
Tried to run calibrate on the Swift-1.5-Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp.gguf model and I got this:
cuda device 0: NVIDIA GeForce RTX 5070 Ti
plan: weights+state+scratch 12845 MB (KV 256 MB for 4096 tokens, prefill 88 MB), expert slots 0 MB, resident experts 0/0
Error: gemm inside batch: x must be quantized already (SAME_X)
Additional logs or context
No response
Activity
Vasyl-Trefilov commentedon Nov 6, 2026
@pberenyi yes, this was my misstake, thanks for bug report! It's working now and it is on main, so just pull new version from main and it should work, please report back, did work yes/no.
You were added in THANKS.md!
Vasyl-Trefilov commentedon Nov 7, 2026
@TheBrutalX This is insane, I will take a look and maybe merge it in main
pberenyi commentedon Nov 7, 2026
Works now, thanks for fixing. Decode speeds are quite decent, up to 93 tok/s.
Vasyl-Trefilov commentedon Nov 7, 2026
@pberenyi No problem, but Strata is not for dense models like "Qwen3.8-27B" the biggest win you will get when you run MoE model like QFN or 35b-a3b, so think about it, maybe give it a shot, t/s there is insane
Vasyl-Trefilov commentedon Nov 8, 2026
This was fixed by #9 (the DeltaNet gate is now prequantized before the batched GEMM, so calibration no longer hits
SAME_X), and @pberenyi confirmed calibration works on main. Closing; reopen with the full output if it comes back.[#10] Warp reduce without __reduce_*_sync (M1-T2) (#11)
docs: close M1-T2 (issue #10, PR #11, fe3eb1a)