Instructions to use Atomic-Germ/Qwen3.8-Distilled-2B-NPU2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Atomic-Germ/Qwen3.8-Distilled-2B-NPU2 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="Atomic-Germ/Qwen3.8-Distilled-2B-NPU2")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Atomic-Germ/Qwen3.8-Distilled-2B-NPU2", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use Atomic-Germ/Qwen3.8-Distilled-2B-NPU2 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 Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0
# Run inference directly in the terminal:
llama cli -hf Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0
# Run inference directly in the terminal:
llama cli -hf Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0

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 Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0
# Run inference directly in the terminal:
./llama-cli -hf Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0

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 Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0

Use Docker

docker model run hf.co/Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0

LM Studio
Jan

vLLM

How to use Atomic-Germ/Qwen3.8-Distilled-2B-NPU2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Atomic-Germ/Qwen3.8-Distilled-2B-NPU2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Atomic-Germ/Qwen3.8-Distilled-2B-NPU2",
		"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/Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0

SGLang

How to use Atomic-Germ/Qwen3.8-Distilled-2B-NPU2 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Atomic-Germ/Qwen3.8-Distilled-2B-NPU2" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Atomic-Germ/Qwen3.8-Distilled-2B-NPU2",
		"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 images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Atomic-Germ/Qwen3.8-Distilled-2B-NPU2" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Atomic-Germ/Qwen3.8-Distilled-2B-NPU2",
		"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"
						}
					}
				]
			}
		]
	}'

Ollama
How to use Atomic-Germ/Qwen3.8-Distilled-2B-NPU2 with Ollama:
```
ollama run hf.co/Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0
```
Unsloth Desktop

How to use Atomic-Germ/Qwen3.8-Distilled-2B-NPU2 with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Atomic-Germ/Qwen3.8-Distilled-2B-NPU2 with Docker Model Runner:
```
docker model run hf.co/Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0
```

Lemonade

How to use Atomic-Germ/Qwen3.8-Distilled-2B-NPU2 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0

Run and chat with the model

lemonade run user.Qwen3.8-Distilled-2B-NPU2-Q8_0

List all available models

lemonade list

Hermes Agent

How to use Atomic-Germ/Qwen3.8-Distilled-2B-NPU2 with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0

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 Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Atomic-Germ/Qwen3.8-Distilled-2B-NPU2 with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0

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 "Atomic-Germ/Qwen3.8-Distilled-2B-NPU2:Q8_0" \
  --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"

Atomic-Germ commited on Aug 24

Commit

d64ebda

verified ·

1 Parent(s): b9c5ca5

Upload q4nx-build.log with huggingface_hub

Browse files

Files changed (1) hide show

q4nx-build.log +443 -0

q4nx-build.log ADDED Viewed

	@@ -0,0 +1,443 @@

+[INFO] Base chain: empero-ai/Qwen3.8-2B-Distill-GGUF -> empero-ai/Qwen3.8-2B -> Qwen/Qwen3.5-2B -> Qwen/Qwen3.5-2B-Base
+[INFO] Skeleton source: Atomic-Germ/Qwen3.5-2B-NPU2
+[INFO] Weights type: language
+[INFO] Found GGUF in empero-ai/Qwen3.8-2B-Distill-GGUF: Qwen3.8-2B-Q8_0.gguf
+[INFO] Converting /home/atomic-germ/.cache/huggingface/hub/models--empero-ai--Qwen3.8-2B-Distill-GGUF/snapshots/f4f73582d0b149595450c719b9a7521a03894f9c/Qwen3.8-2B-Q8_0.gguf to /home/atomic-germ/Projects/q4nx-build-v0.2.0/Qwen3.8-2B-Distill-NPU2...
+[INFO] Using Qwen35 converter (variant: ['qwen35-2B', 'qwen3.5-2B'])
+[INFO] Loading Q4NX config from /home/atomic-germ/Projects/q4nx-build-v0.2.0/configs/qwen3.5_2b.json
+[INFO] Creating name maps...
+[INFO] Detected 23 layers for pattern 'blk.{bid}.ssm_beta.weight'
+[INFO] Detected 25 layers for pattern 'blk.{bid}.attn_q_norm.weight'
+[INFO] Detected 25 layers for pattern 'blk.{bid}.attn_q.weight'
+[INFO] Detected 23 layers for pattern 'blk.{bid}.ssm_dt.bias'
+[INFO] Detected 25 layers for pattern 'blk.{bid}.ffn_down.weight'
+[INFO] Detected 25 layers for pattern 'blk.{bid}.attn_norm.weight'
+[INFO] Detected 25 layers for pattern 'blk.{bid}.ffn_up.weight'
+[INFO] Detected 25 layers for pattern 'blk.{bid}.attn_k_norm.weight'
+[INFO] Detected 25 layers for pattern 'blk.{bid}.post_attention_norm.weight'
+[INFO] Detected 25 layers for pattern 'blk.{bid}.attn_output.weight'
+[INFO] Detected 23 layers for pattern 'blk.{bid}.ssm_a'
+[INFO] Detected 23 layers for pattern 'blk.{bid}.ssm_norm.weight'
+[INFO] Detected 25 layers for pattern 'blk.{bid}.attn_v.weight'
+[INFO] Detected 23 layers for pattern 'blk.{bid}.attn_gate.weight'
+[INFO] Detected 23 layers for pattern 'blk.{bid}.attn_qkv.weight'
+[INFO] Detected 25 layers for pattern 'blk.{bid}.ffn_gate.weight'
+[INFO] Detected 25 layers for pattern 'blk.{bid}.attn_k.weight'
+[INFO] Detected 23 layers for pattern 'blk.{bid}.ssm_out.weight'
+[INFO] Detected 23 layers for pattern 'blk.{bid}.ssm_conv1d.weight'
+[INFO] Detected 23 layers for pattern 'blk.{bid}.ssm_alpha.weight'
+	Converted token_embd.weight to model.embed_tokens.weight
+	Converted blk.0.attn_q.weight to model.layers.0.self_attn.q_proj.weight
+	Converted blk.0.attn_q_norm.weight to model.layers.0.self_attn.q_norm.weight
+	Converted blk.0.attn_k.weight to model.layers.0.self_attn.k_proj.weight
+	Converted blk.0.attn_k_norm.weight to model.layers.0.self_attn.k_norm.weight
+	Converted blk.0.attn_v.weight to model.layers.0.self_attn.v_proj.weight
+	Converted blk.0.attn_output.weight to model.layers.0.self_attn.o_proj.weight
+	Converted blk.0.ffn_up.weight to model.layers.0.mlp.up_proj.weight
+	Converted blk.0.ffn_gate.weight to model.layers.0.mlp.gate_proj.weight
+	Converted blk.0.ffn_down.weight to model.layers.0.mlp.down_proj.weight
+	Converted blk.0.attn_norm.weight to model.layers.0.input_layernorm.weight
+	Converted blk.0.post_attention_norm.weight to model.layers.0.post_attention_layernorm.weight
+	Converted blk.0.attn_gate.weight to model.layers.0.self_attn.gate_proj.weight
+	Converted blk.0.attn_qkv.weight to model.layers.0.linear_attn.qkv_proj.weight
+	Converted blk.0.ssm_a to model.layers.0.linear_attn.ssm_a
+	Converted blk.0.ssm_alpha.weight to model.layers.0.linear_attn.ssm_alpha_proj.weight
+	Converted blk.0.ssm_beta.weight to model.layers.0.linear_attn.ssm_beta_proj.weight
+	Converted blk.0.ssm_conv1d.weight to model.layers.0.linear_attn.ssm_conv1d.weight
+	Converted blk.0.ssm_norm.weight to model.layers.0.linear_attn.ssm_norm.weight
+	Converted blk.0.ssm_out.weight to model.layers.0.linear_attn.ssm_out_proj.weight
+	Converted blk.0.ssm_dt.bias to model.layers.0.linear_attn.ssm_dt.bias
+	Converted output_norm.weight to model.norm.weight
+	Converted output.weight to lm_head.weight
+	Converted v.position_embd.weight to model.visual.pos_embed.weight
+	Converted v.patch_embd.weight to model.visual.patch_embed.proj.weight
+	Converted v.patch_embd.weight.1 to model.visual.patch_embed.proj.weight.1
+	Converted v.patch_embd.bias to model.visual.patch_embed.proj.bias
+	Converted v.post_ln.weight to model.visual.merger.norm.weight
+	Converted v.post_ln.bias to model.visual.merger.norm.bias
+	Converted mm.0.weight to model.visual.merger.linear_fc1.weight
+	Converted mm.0.bias to model.visual.merger.linear_fc1.bias
+	Converted mm.2.weight to model.visual.merger.linear_fc2.weight
+	Converted mm.2.bias to model.visual.merger.linear_fc2.bias
+[INFO] Model does not have a lm_head, use embedding weights as lm_head
+Processing tensor: output_norm.weight with type F32 -> model.norm.weight with dtype F32
+Processing tensor: token_embd.weight with type Q8_0 -> model.embed_tokens.weight with dtype Q4_1
+Processing tensor: blk.0.attn_gate.weight with type Q8_0 -> model.layers.0.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.0.attn_norm.weight with type F32 -> model.layers.0.input_layernorm.weight with dtype F32
+Processing tensor: blk.0.attn_qkv.weight with type Q8_0 -> model.layers.0.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.0.ffn_down.weight with type Q8_0 -> model.layers.0.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.0.ffn_gate.weight with type Q8_0 -> model.layers.0.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.0.ffn_up.weight with type Q8_0 -> model.layers.0.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.0.post_attention_norm.weight with type F32 -> model.layers.0.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.0.ssm_a with type F32 -> model.layers.0.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.0.ssm_alpha.weight with type Q8_0 -> model.layers.0.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.0.ssm_beta.weight with type Q8_0 -> model.layers.0.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.0.ssm_conv1d.weight with type F32 -> model.layers.0.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.0.ssm_dt.bias with type F32 -> model.layers.0.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.0.ssm_norm.weight with type F32 -> model.layers.0.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.0.ssm_out.weight with type Q8_0 -> model.layers.0.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.1.attn_gate.weight with type Q8_0 -> model.layers.1.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.1.attn_norm.weight with type F32 -> model.layers.1.input_layernorm.weight with dtype F32
+Processing tensor: blk.1.attn_qkv.weight with type Q8_0 -> model.layers.1.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.1.ffn_down.weight with type Q8_0 -> model.layers.1.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.1.ffn_gate.weight with type Q8_0 -> model.layers.1.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.1.ffn_up.weight with type Q8_0 -> model.layers.1.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.1.post_attention_norm.weight with type F32 -> model.layers.1.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.1.ssm_a with type F32 -> model.layers.1.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.1.ssm_alpha.weight with type Q8_0 -> model.layers.1.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.1.ssm_beta.weight with type Q8_0 -> model.layers.1.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.1.ssm_conv1d.weight with type F32 -> model.layers.1.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.1.ssm_dt.bias with type F32 -> model.layers.1.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.1.ssm_norm.weight with type F32 -> model.layers.1.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.1.ssm_out.weight with type Q8_0 -> model.layers.1.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.2.attn_gate.weight with type Q8_0 -> model.layers.2.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.2.attn_norm.weight with type F32 -> model.layers.2.input_layernorm.weight with dtype F32
+Processing tensor: blk.2.attn_qkv.weight with type Q8_0 -> model.layers.2.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.2.ffn_down.weight with type Q8_0 -> model.layers.2.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.2.ffn_gate.weight with type Q8_0 -> model.layers.2.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.2.ffn_up.weight with type Q8_0 -> model.layers.2.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.2.post_attention_norm.weight with type F32 -> model.layers.2.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.2.ssm_a with type F32 -> model.layers.2.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.2.ssm_alpha.weight with type Q8_0 -> model.layers.2.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.2.ssm_beta.weight with type Q8_0 -> model.layers.2.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.2.ssm_conv1d.weight with type F32 -> model.layers.2.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.2.ssm_dt.bias with type F32 -> model.layers.2.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.2.ssm_norm.weight with type F32 -> model.layers.2.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.2.ssm_out.weight with type Q8_0 -> model.layers.2.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.3.attn_k.weight with type Q8_0 -> model.layers.3.self_attn.k_proj.weight with dtype Q4_1
+Processing tensor: blk.3.attn_k_norm.weight with type F32 -> model.layers.3.self_attn.k_norm.weight with dtype F32
+Processing tensor: blk.3.attn_norm.weight with type F32 -> model.layers.3.input_layernorm.weight with dtype F32
+Processing tensor: blk.3.attn_output.weight with type Q8_0 -> model.layers.3.self_attn.o_proj.weight with dtype Q4_1
+Processing tensor: blk.3.attn_q.weight with type Q8_0 -> model.layers.3.self_attn.q_proj.weight with dtype Q4_1
+[INFO] Seperate q, gate for q_proj
+Processing tensor: blk.3.attn_q_norm.weight with type F32 -> model.layers.3.self_attn.q_norm.weight with dtype F32
+Processing tensor: blk.3.attn_v.weight with type Q8_0 -> model.layers.3.self_attn.v_proj.weight with dtype Q4_1
+Processing tensor: blk.3.ffn_down.weight with type Q8_0 -> model.layers.3.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.3.ffn_gate.weight with type Q8_0 -> model.layers.3.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.3.ffn_up.weight with type Q8_0 -> model.layers.3.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.3.post_attention_norm.weight with type F32 -> model.layers.3.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.4.attn_gate.weight with type Q8_0 -> model.layers.4.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.4.attn_norm.weight with type F32 -> model.layers.4.input_layernorm.weight with dtype F32
+Processing tensor: blk.4.attn_qkv.weight with type Q8_0 -> model.layers.4.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.4.ffn_down.weight with type Q8_0 -> model.layers.4.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.4.ffn_gate.weight with type Q8_0 -> model.layers.4.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.4.ffn_up.weight with type Q8_0 -> model.layers.4.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.4.post_attention_norm.weight with type F32 -> model.layers.4.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.4.ssm_a with type F32 -> model.layers.4.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.4.ssm_alpha.weight with type Q8_0 -> model.layers.4.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.4.ssm_beta.weight with type Q8_0 -> model.layers.4.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.4.ssm_conv1d.weight with type F32 -> model.layers.4.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.4.ssm_dt.bias with type F32 -> model.layers.4.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.4.ssm_norm.weight with type F32 -> model.layers.4.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.4.ssm_out.weight with type Q8_0 -> model.layers.4.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.5.attn_gate.weight with type Q8_0 -> model.layers.5.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.5.attn_norm.weight with type F32 -> model.layers.5.input_layernorm.weight with dtype F32
+Processing tensor: blk.5.attn_qkv.weight with type Q8_0 -> model.layers.5.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.5.ffn_down.weight with type Q8_0 -> model.layers.5.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.5.ffn_gate.weight with type Q8_0 -> model.layers.5.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.5.ffn_up.weight with type Q8_0 -> model.layers.5.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.5.post_attention_norm.weight with type F32 -> model.layers.5.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.5.ssm_a with type F32 -> model.layers.5.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.5.ssm_alpha.weight with type Q8_0 -> model.layers.5.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.5.ssm_beta.weight with type Q8_0 -> model.layers.5.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.5.ssm_conv1d.weight with type F32 -> model.layers.5.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.5.ssm_dt.bias with type F32 -> model.layers.5.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.5.ssm_norm.weight with type F32 -> model.layers.5.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.5.ssm_out.weight with type Q8_0 -> model.layers.5.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.6.attn_gate.weight with type Q8_0 -> model.layers.6.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.6.attn_norm.weight with type F32 -> model.layers.6.input_layernorm.weight with dtype F32
+Processing tensor: blk.6.attn_qkv.weight with type Q8_0 -> model.layers.6.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.6.ffn_down.weight with type Q8_0 -> model.layers.6.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.6.ffn_gate.weight with type Q8_0 -> model.layers.6.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.6.ffn_up.weight with type Q8_0 -> model.layers.6.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.6.post_attention_norm.weight with type F32 -> model.layers.6.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.6.ssm_a with type F32 -> model.layers.6.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.6.ssm_alpha.weight with type Q8_0 -> model.layers.6.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.6.ssm_beta.weight with type Q8_0 -> model.layers.6.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.6.ssm_conv1d.weight with type F32 -> model.layers.6.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.6.ssm_dt.bias with type F32 -> model.layers.6.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.6.ssm_norm.weight with type F32 -> model.layers.6.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.6.ssm_out.weight with type Q8_0 -> model.layers.6.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.7.attn_k.weight with type Q8_0 -> model.layers.7.self_attn.k_proj.weight with dtype Q4_1
+Processing tensor: blk.7.attn_k_norm.weight with type F32 -> model.layers.7.self_attn.k_norm.weight with dtype F32
+Processing tensor: blk.7.attn_norm.weight with type F32 -> model.layers.7.input_layernorm.weight with dtype F32
+Processing tensor: blk.7.attn_output.weight with type Q8_0 -> model.layers.7.self_attn.o_proj.weight with dtype Q4_1
+Processing tensor: blk.7.attn_q.weight with type Q8_0 -> model.layers.7.self_attn.q_proj.weight with dtype Q4_1
+[INFO] Seperate q, gate for q_proj
+Processing tensor: blk.7.attn_q_norm.weight with type F32 -> model.layers.7.self_attn.q_norm.weight with dtype F32
+Processing tensor: blk.7.attn_v.weight with type Q8_0 -> model.layers.7.self_attn.v_proj.weight with dtype Q4_1
+Processing tensor: blk.7.ffn_down.weight with type Q8_0 -> model.layers.7.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.7.ffn_gate.weight with type Q8_0 -> model.layers.7.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.7.ffn_up.weight with type Q8_0 -> model.layers.7.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.7.post_attention_norm.weight with type F32 -> model.layers.7.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.8.attn_gate.weight with type Q8_0 -> model.layers.8.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.8.attn_norm.weight with type F32 -> model.layers.8.input_layernorm.weight with dtype F32
+Processing tensor: blk.8.attn_qkv.weight with type Q8_0 -> model.layers.8.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.8.ffn_down.weight with type Q8_0 -> model.layers.8.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.8.ffn_gate.weight with type Q8_0 -> model.layers.8.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.8.ffn_up.weight with type Q8_0 -> model.layers.8.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.8.post_attention_norm.weight with type F32 -> model.layers.8.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.8.ssm_a with type F32 -> model.layers.8.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.8.ssm_alpha.weight with type Q8_0 -> model.layers.8.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.8.ssm_beta.weight with type Q8_0 -> model.layers.8.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.8.ssm_conv1d.weight with type F32 -> model.layers.8.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.8.ssm_dt.bias with type F32 -> model.layers.8.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.8.ssm_norm.weight with type F32 -> model.layers.8.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.8.ssm_out.weight with type Q8_0 -> model.layers.8.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.9.attn_gate.weight with type Q8_0 -> model.layers.9.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.9.attn_norm.weight with type F32 -> model.layers.9.input_layernorm.weight with dtype F32
+Processing tensor: blk.9.attn_qkv.weight with type Q8_0 -> model.layers.9.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.9.ffn_down.weight with type Q8_0 -> model.layers.9.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.9.ffn_gate.weight with type Q8_0 -> model.layers.9.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.9.ffn_up.weight with type Q8_0 -> model.layers.9.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.9.post_attention_norm.weight with type F32 -> model.layers.9.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.9.ssm_a with type F32 -> model.layers.9.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.9.ssm_alpha.weight with type Q8_0 -> model.layers.9.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.9.ssm_beta.weight with type Q8_0 -> model.layers.9.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.9.ssm_conv1d.weight with type F32 -> model.layers.9.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.9.ssm_dt.bias with type F32 -> model.layers.9.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.9.ssm_norm.weight with type F32 -> model.layers.9.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.9.ssm_out.weight with type Q8_0 -> model.layers.9.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.10.attn_gate.weight with type Q8_0 -> model.layers.10.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.10.attn_norm.weight with type F32 -> model.layers.10.input_layernorm.weight with dtype F32
+Processing tensor: blk.10.attn_qkv.weight with type Q8_0 -> model.layers.10.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.10.ffn_down.weight with type Q8_0 -> model.layers.10.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.10.ffn_gate.weight with type Q8_0 -> model.layers.10.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.10.ffn_up.weight with type Q8_0 -> model.layers.10.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.10.post_attention_norm.weight with type F32 -> model.layers.10.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.10.ssm_a with type F32 -> model.layers.10.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.10.ssm_alpha.weight with type Q8_0 -> model.layers.10.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.10.ssm_beta.weight with type Q8_0 -> model.layers.10.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.10.ssm_conv1d.weight with type F32 -> model.layers.10.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.10.ssm_dt.bias with type F32 -> model.layers.10.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.10.ssm_norm.weight with type F32 -> model.layers.10.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.10.ssm_out.weight with type Q8_0 -> model.layers.10.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.11.attn_k.weight with type Q8_0 -> model.layers.11.self_attn.k_proj.weight with dtype Q4_1
+Processing tensor: blk.11.attn_k_norm.weight with type F32 -> model.layers.11.self_attn.k_norm.weight with dtype F32
+Processing tensor: blk.11.attn_norm.weight with type F32 -> model.layers.11.input_layernorm.weight with dtype F32
+Processing tensor: blk.11.attn_output.weight with type Q8_0 -> model.layers.11.self_attn.o_proj.weight with dtype Q4_1
+Processing tensor: blk.11.attn_q.weight with type Q8_0 -> model.layers.11.self_attn.q_proj.weight with dtype Q4_1
+[INFO] Seperate q, gate for q_proj
+Processing tensor: blk.11.attn_q_norm.weight with type F32 -> model.layers.11.self_attn.q_norm.weight with dtype F32
+Processing tensor: blk.11.attn_v.weight with type Q8_0 -> model.layers.11.self_attn.v_proj.weight with dtype Q4_1
+Processing tensor: blk.11.ffn_down.weight with type Q8_0 -> model.layers.11.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.11.ffn_gate.weight with type Q8_0 -> model.layers.11.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.11.ffn_up.weight with type Q8_0 -> model.layers.11.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.11.post_attention_norm.weight with type F32 -> model.layers.11.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.12.attn_gate.weight with type Q8_0 -> model.layers.12.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.12.attn_norm.weight with type F32 -> model.layers.12.input_layernorm.weight with dtype F32
+Processing tensor: blk.12.attn_qkv.weight with type Q8_0 -> model.layers.12.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.12.ffn_down.weight with type Q8_0 -> model.layers.12.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.12.ffn_gate.weight with type Q8_0 -> model.layers.12.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.12.ffn_up.weight with type Q8_0 -> model.layers.12.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.12.post_attention_norm.weight with type F32 -> model.layers.12.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.12.ssm_a with type F32 -> model.layers.12.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.12.ssm_alpha.weight with type Q8_0 -> model.layers.12.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.12.ssm_beta.weight with type Q8_0 -> model.layers.12.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.12.ssm_conv1d.weight with type F32 -> model.layers.12.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.12.ssm_dt.bias with type F32 -> model.layers.12.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.12.ssm_norm.weight with type F32 -> model.layers.12.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.12.ssm_out.weight with type Q8_0 -> model.layers.12.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.13.attn_gate.weight with type Q8_0 -> model.layers.13.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.13.attn_norm.weight with type F32 -> model.layers.13.input_layernorm.weight with dtype F32
+Processing tensor: blk.13.attn_qkv.weight with type Q8_0 -> model.layers.13.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.13.ffn_down.weight with type Q8_0 -> model.layers.13.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.13.ffn_gate.weight with type Q8_0 -> model.layers.13.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.13.ffn_up.weight with type Q8_0 -> model.layers.13.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.13.post_attention_norm.weight with type F32 -> model.layers.13.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.13.ssm_a with type F32 -> model.layers.13.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.13.ssm_alpha.weight with type Q8_0 -> model.layers.13.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.13.ssm_beta.weight with type Q8_0 -> model.layers.13.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.13.ssm_conv1d.weight with type F32 -> model.layers.13.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.13.ssm_dt.bias with type F32 -> model.layers.13.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.13.ssm_norm.weight with type F32 -> model.layers.13.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.13.ssm_out.weight with type Q8_0 -> model.layers.13.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.14.attn_gate.weight with type Q8_0 -> model.layers.14.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.14.attn_norm.weight with type F32 -> model.layers.14.input_layernorm.weight with dtype F32
+Processing tensor: blk.14.attn_qkv.weight with type Q8_0 -> model.layers.14.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.14.ffn_down.weight with type Q8_0 -> model.layers.14.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.14.ffn_gate.weight with type Q8_0 -> model.layers.14.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.14.ffn_up.weight with type Q8_0 -> model.layers.14.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.14.post_attention_norm.weight with type F32 -> model.layers.14.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.14.ssm_a with type F32 -> model.layers.14.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.14.ssm_alpha.weight with type Q8_0 -> model.layers.14.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.14.ssm_beta.weight with type Q8_0 -> model.layers.14.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.14.ssm_conv1d.weight with type F32 -> model.layers.14.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.14.ssm_dt.bias with type F32 -> model.layers.14.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.14.ssm_norm.weight with type F32 -> model.layers.14.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.14.ssm_out.weight with type Q8_0 -> model.layers.14.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.15.attn_k.weight with type Q8_0 -> model.layers.15.self_attn.k_proj.weight with dtype Q4_1
+Processing tensor: blk.15.attn_k_norm.weight with type F32 -> model.layers.15.self_attn.k_norm.weight with dtype F32
+Processing tensor: blk.15.attn_norm.weight with type F32 -> model.layers.15.input_layernorm.weight with dtype F32
+Processing tensor: blk.15.attn_output.weight with type Q8_0 -> model.layers.15.self_attn.o_proj.weight with dtype Q4_1
+Processing tensor: blk.15.attn_q.weight with type Q8_0 -> model.layers.15.self_attn.q_proj.weight with dtype Q4_1
+[INFO] Seperate q, gate for q_proj
+Processing tensor: blk.15.attn_q_norm.weight with type F32 -> model.layers.15.self_attn.q_norm.weight with dtype F32
+Processing tensor: blk.15.attn_v.weight with type Q8_0 -> model.layers.15.self_attn.v_proj.weight with dtype Q4_1
+Processing tensor: blk.15.ffn_down.weight with type Q8_0 -> model.layers.15.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.15.ffn_gate.weight with type Q8_0 -> model.layers.15.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.15.ffn_up.weight with type Q8_0 -> model.layers.15.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.15.post_attention_norm.weight with type F32 -> model.layers.15.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.16.attn_gate.weight with type Q8_0 -> model.layers.16.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.16.attn_norm.weight with type F32 -> model.layers.16.input_layernorm.weight with dtype F32
+Processing tensor: blk.16.attn_qkv.weight with type Q8_0 -> model.layers.16.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.16.ffn_down.weight with type Q8_0 -> model.layers.16.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.16.ffn_gate.weight with type Q8_0 -> model.layers.16.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.16.ffn_up.weight with type Q8_0 -> model.layers.16.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.16.post_attention_norm.weight with type F32 -> model.layers.16.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.16.ssm_a with type F32 -> model.layers.16.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.16.ssm_alpha.weight with type Q8_0 -> model.layers.16.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.16.ssm_beta.weight with type Q8_0 -> model.layers.16.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.16.ssm_conv1d.weight with type F32 -> model.layers.16.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.16.ssm_dt.bias with type F32 -> model.layers.16.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.16.ssm_norm.weight with type F32 -> model.layers.16.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.16.ssm_out.weight with type Q8_0 -> model.layers.16.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.17.attn_gate.weight with type Q8_0 -> model.layers.17.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.17.attn_norm.weight with type F32 -> model.layers.17.input_layernorm.weight with dtype F32
+Processing tensor: blk.17.attn_qkv.weight with type Q8_0 -> model.layers.17.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.17.ffn_down.weight with type Q8_0 -> model.layers.17.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.17.ffn_gate.weight with type Q8_0 -> model.layers.17.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.17.ffn_up.weight with type Q8_0 -> model.layers.17.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.17.post_attention_norm.weight with type F32 -> model.layers.17.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.17.ssm_a with type F32 -> model.layers.17.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.17.ssm_alpha.weight with type Q8_0 -> model.layers.17.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.17.ssm_beta.weight with type Q8_0 -> model.layers.17.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.17.ssm_conv1d.weight with type F32 -> model.layers.17.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.17.ssm_dt.bias with type F32 -> model.layers.17.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.17.ssm_norm.weight with type F32 -> model.layers.17.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.17.ssm_out.weight with type Q8_0 -> model.layers.17.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.18.attn_gate.weight with type Q8_0 -> model.layers.18.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.18.attn_norm.weight with type F32 -> model.layers.18.input_layernorm.weight with dtype F32
+Processing tensor: blk.18.attn_qkv.weight with type Q8_0 -> model.layers.18.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.18.ffn_down.weight with type Q8_0 -> model.layers.18.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.18.ffn_gate.weight with type Q8_0 -> model.layers.18.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.18.ffn_up.weight with type Q8_0 -> model.layers.18.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.18.post_attention_norm.weight with type F32 -> model.layers.18.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.18.ssm_a with type F32 -> model.layers.18.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.18.ssm_alpha.weight with type Q8_0 -> model.layers.18.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.18.ssm_beta.weight with type Q8_0 -> model.layers.18.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.18.ssm_conv1d.weight with type F32 -> model.layers.18.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.18.ssm_dt.bias with type F32 -> model.layers.18.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.18.ssm_norm.weight with type F32 -> model.layers.18.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.18.ssm_out.weight with type Q8_0 -> model.layers.18.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.19.attn_k.weight with type Q8_0 -> model.layers.19.self_attn.k_proj.weight with dtype Q4_1
+Processing tensor: blk.19.attn_k_norm.weight with type F32 -> model.layers.19.self_attn.k_norm.weight with dtype F32
+Processing tensor: blk.19.attn_norm.weight with type F32 -> model.layers.19.input_layernorm.weight with dtype F32
+Processing tensor: blk.19.attn_output.weight with type Q8_0 -> model.layers.19.self_attn.o_proj.weight with dtype Q4_1
+Processing tensor: blk.19.attn_q.weight with type Q8_0 -> model.layers.19.self_attn.q_proj.weight with dtype Q4_1
+[INFO] Seperate q, gate for q_proj
+Processing tensor: blk.19.attn_q_norm.weight with type F32 -> model.layers.19.self_attn.q_norm.weight with dtype F32
+Processing tensor: blk.19.attn_v.weight with type Q8_0 -> model.layers.19.self_attn.v_proj.weight with dtype Q4_1
+Processing tensor: blk.19.ffn_down.weight with type Q8_0 -> model.layers.19.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.19.ffn_gate.weight with type Q8_0 -> model.layers.19.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.19.ffn_up.weight with type Q8_0 -> model.layers.19.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.19.post_attention_norm.weight with type F32 -> model.layers.19.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.20.attn_gate.weight with type Q8_0 -> model.layers.20.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.20.attn_norm.weight with type F32 -> model.layers.20.input_layernorm.weight with dtype F32
+Processing tensor: blk.20.attn_qkv.weight with type Q8_0 -> model.layers.20.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.20.ffn_down.weight with type Q8_0 -> model.layers.20.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.20.ffn_gate.weight with type Q8_0 -> model.layers.20.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.20.ffn_up.weight with type Q8_0 -> model.layers.20.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.20.post_attention_norm.weight with type F32 -> model.layers.20.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.20.ssm_a with type F32 -> model.layers.20.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.20.ssm_alpha.weight with type Q8_0 -> model.layers.20.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.20.ssm_beta.weight with type Q8_0 -> model.layers.20.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.20.ssm_conv1d.weight with type F32 -> model.layers.20.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.20.ssm_dt.bias with type F32 -> model.layers.20.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.20.ssm_norm.weight with type F32 -> model.layers.20.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.20.ssm_out.weight with type Q8_0 -> model.layers.20.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.21.attn_gate.weight with type Q8_0 -> model.layers.21.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.21.attn_norm.weight with type F32 -> model.layers.21.input_layernorm.weight with dtype F32
+Processing tensor: blk.21.attn_qkv.weight with type Q8_0 -> model.layers.21.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.21.ffn_down.weight with type Q8_0 -> model.layers.21.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.21.ffn_gate.weight with type Q8_0 -> model.layers.21.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.21.ffn_up.weight with type Q8_0 -> model.layers.21.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.21.post_attention_norm.weight with type F32 -> model.layers.21.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.21.ssm_a with type F32 -> model.layers.21.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.21.ssm_alpha.weight with type Q8_0 -> model.layers.21.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.21.ssm_beta.weight with type Q8_0 -> model.layers.21.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.21.ssm_conv1d.weight with type F32 -> model.layers.21.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.21.ssm_dt.bias with type F32 -> model.layers.21.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.21.ssm_norm.weight with type F32 -> model.layers.21.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.21.ssm_out.weight with type Q8_0 -> model.layers.21.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.22.attn_gate.weight with type Q8_0 -> model.layers.22.self_attn.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.22.attn_norm.weight with type F32 -> model.layers.22.input_layernorm.weight with dtype F32
+Processing tensor: blk.22.attn_qkv.weight with type Q8_0 -> model.layers.22.linear_attn.qkv_proj.weight with dtype Q4_1
+Processing tensor: blk.22.ffn_down.weight with type Q8_0 -> model.layers.22.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.22.ffn_gate.weight with type Q8_0 -> model.layers.22.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.22.ffn_up.weight with type Q8_0 -> model.layers.22.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.22.post_attention_norm.weight with type F32 -> model.layers.22.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.22.ssm_a with type F32 -> model.layers.22.linear_attn.ssm_a with dtype F32
+Processing tensor: blk.22.ssm_alpha.weight with type Q8_0 -> model.layers.22.linear_attn.ssm_alpha_proj.weight with dtype Q8_0
+Processing tensor: blk.22.ssm_beta.weight with type Q8_0 -> model.layers.22.linear_attn.ssm_beta_proj.weight with dtype Q8_0
+Processing tensor: blk.22.ssm_conv1d.weight with type F32 -> model.layers.22.linear_attn.ssm_conv1d.weight with dtype F32
+[INFO] transpose conv1d
+Processing tensor: blk.22.ssm_dt.bias with type F32 -> model.layers.22.linear_attn.ssm_dt.bias with dtype F32
+Processing tensor: blk.22.ssm_norm.weight with type F32 -> model.layers.22.linear_attn.ssm_norm.weight with dtype F32
+Processing tensor: blk.22.ssm_out.weight with type Q8_0 -> model.layers.22.linear_attn.ssm_out_proj.weight with dtype Q8_0
+Processing tensor: blk.23.attn_k.weight with type Q8_0 -> model.layers.23.self_attn.k_proj.weight with dtype Q4_1
+Processing tensor: blk.23.attn_k_norm.weight with type F32 -> model.layers.23.self_attn.k_norm.weight with dtype F32
+Processing tensor: blk.23.attn_norm.weight with type F32 -> model.layers.23.input_layernorm.weight with dtype F32
+Processing tensor: blk.23.attn_output.weight with type Q8_0 -> model.layers.23.self_attn.o_proj.weight with dtype Q4_1
+Processing tensor: blk.23.attn_q.weight with type Q8_0 -> model.layers.23.self_attn.q_proj.weight with dtype Q4_1
+[INFO] Seperate q, gate for q_proj
+Processing tensor: blk.23.attn_q_norm.weight with type F32 -> model.layers.23.self_attn.q_norm.weight with dtype F32
+Processing tensor: blk.23.attn_v.weight with type Q8_0 -> model.layers.23.self_attn.v_proj.weight with dtype Q4_1
+Processing tensor: blk.23.ffn_down.weight with type Q8_0 -> model.layers.23.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.23.ffn_gate.weight with type Q8_0 -> model.layers.23.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.23.ffn_up.weight with type Q8_0 -> model.layers.23.mlp.up_proj.weight with dtype Q4_1
+Processing tensor: blk.23.post_attention_norm.weight with type F32 -> model.layers.23.post_attention_layernorm.weight with dtype F32
+Processing tensor: blk.24.attn_k.weight with type Q8_0 -> model.layers.24.self_attn.k_proj.weight with dtype Q4_1
+Processing tensor: blk.24.attn_k_norm.weight with type F32 -> model.layers.24.self_attn.k_norm.weight with dtype F32
+Processing tensor: blk.24.attn_norm.weight with type F32 -> model.layers.24.input_layernorm.weight with dtype F32
+Processing tensor: blk.24.attn_output.weight with type Q8_0 -> model.layers.24.self_attn.o_proj.weight with dtype Q4_1
+Processing tensor: blk.24.attn_q.weight with type Q8_0 -> model.layers.24.self_attn.q_proj.weight with dtype Q4_1
+[INFO] Seperate q, gate for q_proj
+Processing tensor: blk.24.attn_q_norm.weight with type F32 -> model.layers.24.self_attn.q_norm.weight with dtype F32
+Processing tensor: blk.24.attn_v.weight with type Q8_0 -> model.layers.24.self_attn.v_proj.weight with dtype Q4_1
+Processing tensor: blk.24.ffn_down.weight with type Q8_0 -> model.layers.24.mlp.down_proj.weight with dtype Q4_1
+Processing tensor: blk.24.ffn_gate.weight with type Q8_0 -> model.layers.24.mlp.gate_proj.weight with dtype Q4_1
+Processing tensor: blk.24.ffn_up.weight with type Q8_0 -> model.layers.24.mlp.up_proj.weight with dtype Q4_1
+[SKIP] blk.24.nextn.eh_proj.weight (MTP next-token prediction weights, absent from official Q4NX)
+[SKIP] blk.24.nextn.enorm.weight (MTP next-token prediction weights, absent from official Q4NX)
+[SKIP] blk.24.nextn.hnorm.weight (MTP next-token prediction weights, absent from official Q4NX)
+[SKIP] blk.24.nextn.shared_head_norm.weight (MTP next-token prediction weights, absent from official Q4NX)
+Processing tensor: blk.24.post_attention_norm.weight with type F32 -> model.layers.24.post_attention_layernorm.weight with dtype F32
+[INFO] Extracting tokenizer JSON...
+[INFO] EOS token ID: 248046
+[INFO] Padding token ID: 248044
+[INFO] Tokenizer saved to /home/atomic-germ/Projects/q4nx-build-v0.2.0/Qwen3.8-2B-Distill-NPU2/tokenizer.json
+[INFO] Saving Q4NX tensors to /home/atomic-germ/Projects/q4nx-build-v0.2.0/Qwen3.8-2B-Distill-NPU2/model.q4nx...
+[INFO] Copied model assets from local source: /home/atomic-germ/.cache/huggingface/hub/models--Atomic-Germ--Qwen3.5-2B-NPU2/snapshots/92b00a5dc3487149c471c158628f63d9a621bb25
+[INFO] Model assets present: ['vision_weight.q4nx', 'config.json', 'tokenizer.json', 'tokenizer_config.json', 'chat_template.jinja']
+[INFO] Keeping skeleton-provided vision_config (matches shipped vision weights)
+[INFO] Patched tokenizer_config.json with token ids from GGUF metadata
+[INFO] Writing README.md based on Atomic-Germ/Qwen3.5-2B-NPU2's model card
+[INFO] Model directory ready: /home/atomic-germ/Projects/q4nx-build-v0.2.0/Qwen3.8-2B-Distill-NPU2
+[WARN] Overwriting existing model directory: /home/atomic-germ/.config/flm/models/Qwen3.8-Distill-2B-NPU2
+[INFO] Deployed model files to: /home/atomic-germ/.config/flm/models/Qwen3.8-Distill-2B-NPU2
+[INFO] Registered tag 'qwen3.8-distill:2b' in: /home/atomic-germ/.config/flm/model_list.json
+[INFO] Registry defaults copied from official entry: qwen3.5:2b
+[INFO] FLM does not read this registry yet. Add to your shell rc:
+        export FLM_CONFIG_PATH=/home/atomic-germ/.config/flm/model_list.json
+[INFO] Linked xclbins for 'Qwen3.8-Distill-2B-NPU2' -> 'Qwen3.5-2B-NPU2'
+[INFO] The runtime needs to find these xclbins. Add to your shell rc:
+        export FLM_XCLBIN_PATH=/home/atomic-germ/.config/flm
+[INFO] Conversion complete! Output saved to /home/atomic-germ/Projects/q4nx-build-v0.2.0/Qwen3.8-2B-Distill-NPU2