Instructions to use vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16
# Run inference directly in the terminal:
llama cli -hf vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16
# Run inference directly in the terminal:
llama cli -hf vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16
# Run inference directly in the terminal:
./llama-cli -hf vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16

Use Docker

docker model run hf.co/vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16

LM Studio
Jan

vLLM

How to use vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16

Ollama
How to use vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF with Ollama:
```
ollama run hf.co/vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16
```

Unsloth Studio

How to use vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF to start chatting

How to use vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF with Docker Model Runner:
```
docker model run hf.co/vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16
```

Lemonade

How to use vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16

Run and chat with the model

lemonade run user.Muse-Glimmer-30B-ROCmFPX-GGUF-BF16

List all available models

lemonade list

Hermes Agent

How to use vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF:BF16" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

vmlinux commited on 8 days ago

Commit

9796a93

verified ·

1 Parent(s): ddf227b

Document DFlash conversion and ROCm benchmarks

Browse files

Files changed (2) hide show

BUILD_PLAN.md +25 -0
BUILD_RESULTS.md +33 -0

BUILD_PLAN.md CHANGED Viewed

@@ -4,6 +4,8 @@
 - Hugging Face model: `meta-models/Muse-Glimmer-30B`
 - Model revision: `f84ecc3a0ea984a4c04542a84269e3d065350a6e`
 - ROCmFPX base: `00d54526e24e3aba4c76474e3147cbf9c7cc034c`
 - Upstream Muse support: `62bf73d25c53b8161f8a22894d4f90c4aebbd7d0`
 - Upstream converter worktree: `d2f83055dca6dd009d8a52bdff792fbb286f4444`
@@ -21,6 +23,8 @@ canonical `/home/vmlinux/models/llama.cpp` checkout or build.
 | ROCmFP4 | `Q4_0_ROCMFP4_STRIX` | Strix Halo speed/quality recipe; dual-scale FP4 attention K/V and Q6_K token embeddings |
 | ROCmFP8 | `Q8_0_ROCMFPX` | High-fidelity 8.25-bpw reference layout |
 | ROCmFP4-Q6-QUALITY | `Q4_0_ROCMFP4_COHERENT` | Dual-scale FP4 body with Q6_K token embeddings for the coherence-biased quality build |
 The two FP4 builds use a 500-chunk importance matrix generated from the local
 `narrative-general-imatrix-sample.txt` corpus. This supplies varied prose,
@@ -29,9 +33,14 @@ provides the agent/tool-oriented tensor routing. The FP8 reference does not
 need an importance matrix because its preset does not consume importance
 weights.
 ## Paths
 - Source snapshot: `source/Muse-Glimmer-30B/`
 - BF16 text model and projector: `intermediate/`
 - Importance matrix: `calibration/`
 - Final models: `artifacts/`
@@ -61,6 +70,22 @@ toolchain/build-muse-rocmfpx/bin/llama-quantize \
   intermediate/Muse-Glimmer-30B-BF16.gguf \
   artifacts/Muse-Glimmer-30B-ROCmFP4-Q6-QUALITY.gguf \
   Q4_0_ROCMFP4_COHERENT
 ```
 Exact tensor inventories, sizes, checksums, and smoke-test results are in

 - Hugging Face model: `meta-models/Muse-Glimmer-30B`
 - Model revision: `f84ecc3a0ea984a4c04542a84269e3d065350a6e`
+- Hugging Face drafter: `meta-models/Muse-Glimmer-30B-assistant`
+- Drafter revision: `2c86316d689027b91123638739743fef1d425233`
 - ROCmFPX base: `00d54526e24e3aba4c76474e3147cbf9c7cc034c`
 - Upstream Muse support: `62bf73d25c53b8161f8a22894d4f90c4aebbd7d0`
 - Upstream converter worktree: `d2f83055dca6dd009d8a52bdff792fbb286f4444`
 | ROCmFP4 | `Q4_0_ROCMFP4_STRIX` | Strix Halo speed/quality recipe; dual-scale FP4 attention K/V and Q6_K token embeddings |
 | ROCmFP8 | `Q8_0_ROCMFPX` | High-fidelity 8.25-bpw reference layout |
 | ROCmFP4-Q6-QUALITY | `Q4_0_ROCMFP4_COHERENT` | Dual-scale FP4 body with Q6_K token embeddings for the coherence-biased quality build |
+| DFlash ROCmFP4 | `Q4_0_ROCMFP4_STRIX` | Small, speed-oriented official DFlash companion; recommended on Strix Halo |
+| DFlash ROCmFP8 | `Q8_0_ROCMFPX` | Higher-precision DFlash reference for acceptance and hardware tuning |
 The two FP4 builds use a 500-chunk importance matrix generated from the local
 `narrative-general-imatrix-sample.txt` corpus. This supplies varied prose,
 need an importance matrix because its preset does not consume importance
 weights.
+The DFlash drafters do not use the main-model iMatrix. They are quantized from
+the official assistant checkpoint, and the FP4 conversion keeps the feature
+fusion `fc.weight` at Q8_0 as a precision floor.
 ## Paths
 - Source snapshot: `source/Muse-Glimmer-30B/`
+- DFlash source snapshot: `source/Muse-Glimmer-30B-assistant/`
 - BF16 text model and projector: `intermediate/`
 - Importance matrix: `calibration/`
 - Final models: `artifacts/`
   intermediate/Muse-Glimmer-30B-BF16.gguf \
   artifacts/Muse-Glimmer-30B-ROCmFP4-Q6-QUALITY.gguf \
   Q4_0_ROCMFP4_COHERENT
+.venv/bin/python toolchain/scripts/convert_hf_to_gguf_modular.py \
+  source/Muse-Glimmer-30B-assistant \
+  --target-model-dir source/Muse-Glimmer-30B \
+  --outtype bf16 \
+  --outfile intermediate/Muse-Glimmer-30B-DFlash-BF16.gguf
+toolchain/build-muse-rocmfpx/bin/llama-quantize \
+  intermediate/Muse-Glimmer-30B-DFlash-BF16.gguf \
+  artifacts/Muse-Glimmer-30B-DFlash-ROCmFP4.gguf \
+  Q4_0_ROCMFP4_STRIX
+toolchain/build-muse-rocmfpx/bin/llama-quantize \
+  intermediate/Muse-Glimmer-30B-DFlash-BF16.gguf \
+  artifacts/Muse-Glimmer-30B-DFlash-ROCmFP8.gguf \
+  Q8_0_ROCMFPX
 ```
 Exact tensor inventories, sizes, checksums, and smoke-test results are in

BUILD_RESULTS.md CHANGED Viewed

@@ -8,11 +8,19 @@
 | `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 | — |
 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.
 ## Importance matrix
 - File: `calibration/Muse-Glimmer-30B-narrative-general.imatrix`
@@ -27,6 +35,10 @@ 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.
 ## Runtime validation
 All checks used the experimental Muse-capable ROCmFPX build on `ROCm0` with
@@ -39,6 +51,25 @@ all layers offloaded and a 1,024-token context unless noted.
 | 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 |
 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
@@ -51,6 +82,8 @@ at upload time.
 - Vulkan ROCmFP4 copy/conversion backend tests: 34/34 passed
 - `llama-quantize`, `llama-imatrix`, `llama-cli`, and `llama-server`: built
 - ROCmFPX build version: `209 (00d54526e)`
 SHA-256 digests for sources, intermediates, and final outputs are recorded in
 `manifests/SHA256SUMS`.

 | `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`
 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
 | 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 decode | Speedup | Range | Weighted acceptance |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| Baseline | — | 13.7 tok/s | 1.00× | 13.7–13.7 | — |
+| DFlash ROCmFP4 | 6 | 28.3 tok/s | 2.07× | 24.3–31.8 | 519/1,448 (35.8%) |
+| DFlash ROCmFP8 | 6 | 27.2 tok/s | 1.99× | 24.1–33.0 | 519/1,454 (35.7%) |
+| DFlash ROCmFP4 | 15 | 24.6 tok/s | 1.80× | 18.5–30.7 | 543/3,171 (17.1%) |
+| DFlash ROCmFP8 | 15 | 26.3 tok/s | 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
 - 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`.