# Build results

## Outputs

| File | Preset / type | Bytes | Size | BPW |
| --- | --- | ---: | ---: | ---: |
| `artifacts/Muse-Glimmer-30B-ROCmFP4.gguf` | `Q4_0_ROCMFP4_STRIX`, ftype 105 | 15,210,123,424 | 14.17 GiB | 4.36 |
| `artifacts/Muse-Glimmer-30B-ROCmFP8.gguf` | `Q8_0_ROCMFPX`, ftype 111 | 28,742,539,776 | 26.77 GiB | 8.25 |
| `artifacts/Muse-Glimmer-30B-ROCmFP4-Q6-QUALITY.gguf` | `Q4_0_ROCMFP4_COHERENT`, ftype 102 | 16,032,978,080 | 14.94 GiB | 4.60 |
| `intermediate/mmproj-Muse-Glimmer-30B-BF16.gguf` | BF16 vision projector | 3,849,173,920 | 3.59 GiB | — |
| `artifacts/Muse-Glimmer-30B-DFlash-ROCmFP4.gguf` | `Q4_0_ROCMFP4_STRIX` drafter | 1,491,030,176 | 1.39 GiB | 4.63 |
| `artifacts/Muse-Glimmer-30B-DFlash-ROCmFP8.gguf` | `Q8_0_ROCMFPX` drafter | 2,649,174,176 | 2.47 GiB | 8.25 |

The text GGUFs each contain 731 tensors and report the `muse-glimmer`
architecture. The projector contains 809 tensors, 50 vision blocks, and the
`muse-glimmer` projector type.

The DFlash GGUFs each contain 58 tensors and report the `dflash` architecture.
Their source is Meta's official five-layer `MuseGlimmerAssistantModel` at
revision `2c86316d689027b91123638739743fef1d425233`. The modular ROCmFPX
conversion and current upstream conversion produced byte-identical BF16 GGUFs
(`dc4ecbb…ad891`) before quantization.

## Importance matrix

- File: `calibration/Muse-Glimmer-30B-narrative-general.imatrix`
- Format: GGUF importance matrix
- Entries consumed by quantizer: 416
- Calibration: 500 chunks × 512 tokens (approximately 256k tokens)
- Input corpus: local `narrative-general-imatrix-sample.txt`
- Recovery files: `.at_100`, `.at_200`, `.at_300`, `.at_400`, `.at_500`

Both FP4 quantizers parsed all 416 entries and reported that the matrix was
computed from 500 chunks. The standalone `--show-statistics` mode exits with
SIGFPE in this ROCmFPX base; this does not affect matrix generation or the two
successful quantizer consumers.

Neither DFlash drafter uses an iMatrix. They were quantized directly from the
official BF16 assistant; the custom FP4 precision floor retains `fc.weight` as
Q8_0 while routing the remaining eligible weights through the Strix FP4 preset.

## Runtime validation

All checks used the experimental Muse-capable ROCmFPX build on `ROCm0` with
all layers offloaded and a 1,024-token context unless noted.

| Model | Result | Prompt rate | Decode rate |
| --- | --- | ---: | ---: |
| ROCmFP4 | clean 16-token single-turn generation | 113.7 tok/s | 14.9 tok/s |
| ROCmFP8 | clean 16-token single-turn generation | 96.7 tok/s | 7.8 tok/s |
| ROCmFP4-Q6-QUALITY | clean 16-token single-turn generation | 39.0 tok/s | 14.0 tok/s |
| ROCmFP4 + BF16 mmproj | loaded `text, vision`, decoded a PNG as a folder icon | 81.7 tok/s | 14.9 tok/s |

### DFlash throughput

Controlled DFlash runs used the ROCmFP4 target, three prompts, 256 generated
tokens per prompt, context 2,048, batch size 1, greedy decoding, flash
attention, full ROCm0 offload, and Q4_0 KV caches for both contexts.

| Mode | Draft window | Mean prompt t/s | Mean output t/s | Speedup | Output range | Weighted acceptance |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| Baseline | — | 73.7 | 13.7 | 1.00× | 13.7–13.7 | — |
| DFlash ROCmFP4 | 6 | 65.5 | 28.3 | 2.07× | 24.3–31.8 | 519/1,448 (35.8%) |
| DFlash ROCmFP8 | 6 | 65.1 | 27.2 | 1.99× | 24.1–33.0 | 519/1,454 (35.7%) |
| DFlash ROCmFP4 | 15 | 65.6 | 24.6 | 1.80× | 18.5–30.7 | 543/3,171 (17.1%) |
| DFlash ROCmFP8 | 15 | 65.3 | 26.3 | 1.92× | 19.2–34.6 | 563/2,891 (19.5%) |

The six-token window is the recommended starting point on this host. The
runtime patch includes the upstream DFlash injected-cache rotation fix; before
that fix, quantized KV caches produced near-zero acceptance despite valid
target verification.

The vision smoke test used a 2,048-token context and a local system icon. The
older ROCmFPX flash-attention ABI requires the Muse sparse attention mask in
FP16; the experiment toolchain converts the upstream FP32 mask values to FP16
at upload time.

## Toolchain verification

- `test-llama-archs`: passed, including Muse Glimmer architecture loading
- `test-quantize-fns`: passed
- Vulkan ROCmFP4 copy/conversion backend tests: 34/34 passed
- `llama-quantize`, `llama-imatrix`, `llama-cli`, and `llama-server`: built
- ROCmFPX build version: `209 (00d54526e)`
- DFlash ROCmFP4 and ROCmFP8 both completed verified speculative generation
  with Q4_0 KV caches after the injected-cache rotation backport

SHA-256 digests for sources, intermediates, and final outputs are recorded in
`manifests/SHA256SUMS`.