Bug Report

Priority: P1
Reporter: @NigelTufnel12345 (via X)
Status: Confirmed crash, awaiting logs

Environment

• Hardware: M3 Ultra, 512GB RAM
• Model: GPT-OSS-120B, MXFP4 quantization
• Config: -ctk q8_0 -ctv turbo4 -fa on
• Branch: feature/turboquant-kv-cache

Symptom

Immediate segfault on startup. Backtrace fires on:

ggml.c:3246: GGML_ASSERT(ggml_can_mul_mat(a, b)) failed

Model runs fine with regular q8_0 KV (no turbo). Crash is specific to the turbo code path.

Probable Cause

ggml_can_mul_mat assertion indicates tensor shape mismatch. MXFP4 is a newer weight quantization format and the turbo KV insertion or flash attention path likely has an unhandled shape interaction with MXFP4 weight tensors.

Possible areas to investigate:

• Turbo SET_ROWS kernel shape assumptions with MXFP4 input tensors
• Flash attention tile/vec kernel dispatch with non-standard weight types
• head_dim detection or zero-padding logic with MXFP4 metadata

Logs Needed

@NigelTufnel12345 — please upload:

1. Full backtrace (the complete crash output, not just the assertion line)
2. Model startup log (the lines showing n_embd, n_head, n_embd_head_k, and any turbo-related log lines)
3. Exact command used to launch the server/cli

Workaround

Use -ctk q8_0 -ctv q8_0 (no turbo compression) until this is resolved.