[dzannotti]

Re-opens #156, which was reverted in #168 together with the upstream sync it depends on.

Stacked on 169 (the re-opened sync). Merge that one first; this PR's diff is #156 only.

---

Stacked on #155, because it needs the GGUF imatrix statistics from upstream ggml-org#14891, which #155 brings in. When #155 is merged and its branch is deleted, GitHub retargets this PR to master. This replaces halo-box#35, which can't be closed there because that repository is archived.

This ports ggml-org#15550 (EAddario, still open upstream) at 839a43500. It adds --target-bpw, --target-size, --state-file and --target-exclude to llama-quantize: for each tensor, it picks the quant type that reaches the target at the lowest estimated error, using the activation statistics in GGUF imatrices.

• Applied as one commit: the PR's diff against its upstream base cb7934c52, without the PR's stray 1k-line uv.lock.
• common/imatrix-loader.*: the PR's .stats / stats_schema loading is merged with imatrix: calculate activation-based statistics for new format (GGUF) imatrices ggml-org/llama.cpp#14891's n_layer_nextn. The schema names and the .stats suffix match what llama-imatrix writes.
• src/llama-quant.cpp: strix's ROCmFPx rules are kept in every tensor category, renamed to the PR's TENSOR_CATEGORY_* enum. The ROCmFP4 / ROCmFPx fallbacks move into the PR's fallback_type_for().
• When upstream merges quantize: add option to automatically choose optimal quant types to reach a file/bpw target size at lowest error ggml-org/llama.cpp#15550 (possibly changed), the next sync will conflict in llama-quant.cpp and quantize.cpp. Take upstream's side there and re-apply the ROCmFPx rules.

Measured

Qwen3-1.7B, requantized from Q8_0 on Vulkan. The imatrix comes from docs/ (60×512 tokens). Perplexity is measured on held-out tools/ + examples/ READMEs (60×512 tokens).

model	size	BPW	PPL
Q4_K_M + imatrix	1217 MiB	5.03	6.4851 ± 0.169
--target-bpw 5.0262	1217 MiB	5.03	6.4884 ± 0.169
Q4_0_ROCMFP4 + imatrix (strix path still works)	1280 MiB	5.28	6.4739 ± 0.168

The target is hit exactly. At Q4_K_M's size the quality is the same within the error bars: one earlier run on the halo port came out slightly ahead, this one is level. Reusing --state-file cuts re-targeting to seconds.

🤖 Generated with Claude Code