GLM-4.7-Flash-REAP-23B-A3B-ROCmFPX-STRIX_LEAN-GGUF

GLM-4.7-Flash-REAP-23B-A3B ROCmFPX STRIX_LEAN — GGUF

ROCmFPX Q4_0_ROCMFP4_STRIX_LEAN quant of cerebras/GLM-4.7-Flash-REAP-23B-A3B (GLM-4.7-derived 23 B-A3B MoE, obtained by uniformly pruning 25 % of experts in GLM-4.7-Flash using the REAP method).

Built with charlie12345/ROCmFPX on a Radeon RX 9060 XT 16 GB (gfx1200), ROCm 7.2.3, NixOS 25.11. Quantized 2026-06-27 with build commit 11d76c2.

This is not a stock llama.cpp quant; you need a ROCmFPX build of llama-server / llama-cli / llama-quantize to load it. Stock llama.cpp will reject the file with unknown quantization.

Scope of these benchmarks — read this first

These numbers are a light baseline, not a thorough ROCmFPX evaluation. The mesh's bench framework is built for production agent workload regression-detection on the local stack, not for the kind of multi-axis sweep that upstream quant maintainers typically publish. Specifically:

• Harness scope is bounded. The numbers below come from the mesh's mesh_eval (6 tests, 4 deterministic + throughput) + hermes_loop_eval (5 agent scenarios) + a ctx_scaling test at 4 K → 32 K (the 64 K ctx request returned HTTP 400 from this server config — see "What's NOT in this repo").
• Sample sizes are small. Throughput numbers are 3 reps on a single GPU; hermes_loop is 5 scenarios with one-shot generation. None are powered for statistical significance on a per-token level.
• No perplexity / wikitext / MMLU / GSM8K. The mesh's stack isn't a quality benchmark — those are upstream ROCmFPX's territory.
• Single GPU class. All measurements are on a 16 GB RDNA4 (RX 9060 XT, gfx1200). No Strix unified-memory, no CDNA, no multi-GPU, no Vulkan, no CUDA. Cross-hardware generalization is not implied.
• No human eval. "Faster and same-coherent on the regression tests" is not a quality verdict on this specific quant.
• Heaviest model in the mesh. GLM REAP 23B at 12 GB is the biggest single-model quant the mesh can serve. On smaller GPUs (<12 GB VRAM), this file will not fit. The 16 GB card runs it with ~3 GB headroom.

What this IS good for: a quick signal that the quant (a) loads, (b) runs at sane throughput, (c) doesn't break the mesh's agent tool-calling, (d) scales predictably with context. What this is NOT good for: claiming "this is the best quant of this model," reproducing academic benchmark results, or substituting for upstream's validation work.

For a rigorous view, the parent repo cerebras/GLM-4.7-Flash-REAP-23B-A3B, the upstream zai-org/GLM-4.7-Flash, and the model's stock GGUF variants (e.g. on unsloth/) are the place to look.

What we measured

Hardware: Node B, AMD Ryzen 9 5900XT 16-core, Radeon RX 9060 XT 16 GB (gfx1200), ROCm 7.2.3, NixOS 25.11 Software: charlie12345/ROCmFPX main @ 11d76c2 Source GGUF: GLM-4.7-Flash-REAP-23B-A3B-BF16.gguf (BF16, 43 GB) — the Unsloth-distributed GGUF of the Cerebras-pruned safetensors Same-stack comparison: Q3_0_ROCMFPX (3-bit ROCmFPX experimental, 12 GB file) on the same source

Agent-loop throughput — STRIX_LEAN vs Q3_0_ROCMFPX (hermes_loop, same harness, same source)

Both quants pass all 5 scenarios. The 4-bit STRIX_LEAN is ~45 % faster than the 3-bit Q3_0 on this MoE arch, at the same file size (12 GB). This is the headline finding for this model.

mesh_eval (raw JSON: raw-mesh-eval-glm-reap-23b-strix-lean.json)

hermes_loop (raw JSON: raw-hermes-loop-glm-reap-23b-strix-lean.json)

Context scaling (raw JSON: ctx-scaling-glm-reap-strix-lean-64k-20260627-143748.json)

Findings:

• Decode throughput holds at 50 t/s across 4 K → 32 K ctx.
• Prompt processing degrades sharply: 4 K → 32 K drops from 669 → 166 pp t/s (4× slower). This is a known property of the GLM-4.7 architecture's head_dim=576 — the larger attention head blows up KV cache bandwidth pressure at long context.
• The 64 K failure is the server's --ctx-size cap, not a model limit. The parent GLM-4.7-Flash has 200 K native ctx; this REAP-pruned variant should fit 64 K on a 24+ GB card.

KV cache type — head_dim=576 constraint (no turbo support)

This model has head_dim=576 (GLM-4.7 architecture). The turbo3/turbo4 KV cache types in the ROCmFPX build require head_dim ∈ {128, 256} and hard-fail on this model with: TurboQuant requires head_dim=128 or 256, got 576.

Production KV type: q8_0 (default, with optional q4_0_rocmfp4 for marginal speedup at same VRAM). See references/rocmfpx-build-quant-bench.md Pattern 13 in the meshina corpus for the full sweep.

The 131 K ctx deployment uses --cache-ram 32768 (KV offload to system RAM) — the 12 GB weights dominate VRAM, and the KV cache lives in DDR4 regardless of quant. This is what makes long-context GLM REAP viable on 16 GB hardware.

Quick start

# Build llama.cpp with ROCmFPX
git clone https://github.com/charlie12345/ROCmFPX
cd ROCmFPX
cmake -S . -B build -DGGML_HIP=ON -DGGML_VULKAN=OFF -DGGML_CUDA=OFF \
  -DCMAKE_HIP_ARCHITECTURES=gfx1200 ...
cmake --build build --target llama-server llama-cli llama-quantize

# Serve (131 072 ctx, q8_0 KV [head_dim=576, turbo incompatible], KV offload, fa=on)
./build/bin/llama-server \
  -m GLM-4.7-Flash-REAP-23B-A3B-ROCmFPX-STRIX_LEAN.gguf \
  -np 1 -c 131072 \
  -ctk q8_0 -ctv q8_0 \
  -kvo -cram 32768 -fa on

Reproduce the quant

SRC=/path/to/GLM-4.7-Flash-REAP-23B-A3B-BF16.gguf

~/ROCmFPX/build-rdna4/bin/llama-quantize \
  "$SRC" \
  GLM-4.7-Flash-REAP-23B-A3B-ROCmFPX-STRIX_LEAN.gguf \
  Q4_0_ROCMFP4_STRIX_LEAN

Quantize time: ~3-5 min warm-cache, CPU-only. Source BF16 is 43 GB so the first cold quant is slower.

Files in this repo

What's NOT in this repo (caveats)

• Stock llama.cpp will not load this file. The ROCmFP4 weight format is unique to charlie12345/ROCmFPX.
• No CUDA / non-AMD GPU bench. All measurements are RDNA4 (gfx1200).
• 64 K ctx is HTTP 400 on this server. The parent GLM-4.7-Flash has 200 K native ctx. Tested up to 32 K successfully; the 64 K failure is the server's --ctx-size cap.
• No turbo3/4 KV cache on this model (head_dim=576). Hard architectural constraint, not a bug.
• The source GGUF is Unsloth-distributed (per general.quantized_by = "Unsloth" in the metadata). The actual safetensors parent is cerebras/GLM-4.7-Flash-REAP-23B-A3B, derived from zai-org/GLM-4.7-Flash (the unpruned 200 K-ctx model). The chain is: safetensors → Unsloth GGUF → our STRIX_LEAN.
• 12 GB minimum VRAM. Doesn't fit on <12 GB cards. The mesh's 16 GB card runs it with ~3 GB headroom.
• No MTP / speculative-decode bench on this file. GLM-4.7 architecture is not MTP-capable in this release.
• No vision/multimodal test. This variant is text-only.
• No quality benchmark (perplexity, MMLU, GSM8K). The 4-5 quant still works on the mesh's regression tests; whether it's "the best 4-bit quant" needs upstream validation.

Provenance

• Source model: cerebras/GLM-4.7-Flash-REAP-23B-A3B — 23 B-A3B MoE, 25 % of experts pruned from zai-org/GLM-4.7-Flash using the REAP method
• Source model license: mit
• Source GGUF uploader: Unsloth (per general.quantized_by in the BF16 source metadata)
• Quantizer: charlie12345/ROCmFPX main @ 11d76c2 (2026-06-27)
• Quantizer license: MIT
• Build hardware: Node B, AMD Ryzen 9 5900XT 16-core, Radeon RX 9060 XT 16 GB (gfx1200), ROCm 7.2.3, NixOS 25.11
• Build tooling: NixOS 25.11, ROCm store paths dynamic-discovered. See the meshina repo's references/nixos-rocm-external-build-recipe.md for the build env setup.
• Bench harnesses: scripts/mesh-bench/mesh_eval.py + scripts/mesh-bench/hermes_loop_eval.py + scripts/mesh-bench/ctx_scaling_bench.py from the meshina repo (private)
• Original bench report: raw/benchmarks/2026-06-27-rocmfpx-validation/briefs/2026-06-27-rocmfpx-rdna4-16gb.md in the meshina repo

License

• The GLM-4.7-Flash-REAP parent is MIT (per its HF model card).
• The charlie12345/ROCmFPX quantizer is MIT.
• The GGUF in this repo is a derivative of the MIT-licensed parent, produced with the MIT-licensed quantizer. The MIT license is preserved.