Instructions to use SC117/Agents-A1-MTP-APEX-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use SC117/Agents-A1-MTP-APEX-GGUF with Transformers:

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

pipe = pipeline("text-generation", model="SC117/Agents-A1-MTP-APEX-GGUF")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("SC117/Agents-A1-MTP-APEX-GGUF", dtype="auto")

llama-cpp-python

How to use SC117/Agents-A1-MTP-APEX-GGUF with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="SC117/Agents-A1-MTP-APEX-GGUF",
	filename="Agents-A1-MTP-APEX-I-Balanced.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use SC117/Agents-A1-MTP-APEX-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 SC117/Agents-A1-MTP-APEX-GGUF:BF16
# Run inference directly in the terminal:
llama cli -hf SC117/Agents-A1-MTP-APEX-GGUF:BF16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf SC117/Agents-A1-MTP-APEX-GGUF:BF16
# Run inference directly in the terminal:
llama cli -hf SC117/Agents-A1-MTP-APEX-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 SC117/Agents-A1-MTP-APEX-GGUF:BF16
# Run inference directly in the terminal:
./llama-cli -hf SC117/Agents-A1-MTP-APEX-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 SC117/Agents-A1-MTP-APEX-GGUF:BF16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf SC117/Agents-A1-MTP-APEX-GGUF:BF16

Use Docker

docker model run hf.co/SC117/Agents-A1-MTP-APEX-GGUF:BF16

LM Studio
Jan

vLLM

How to use SC117/Agents-A1-MTP-APEX-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "SC117/Agents-A1-MTP-APEX-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": "SC117/Agents-A1-MTP-APEX-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/SC117/Agents-A1-MTP-APEX-GGUF:BF16

SGLang

How to use SC117/Agents-A1-MTP-APEX-GGUF 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 "SC117/Agents-A1-MTP-APEX-GGUF" \
    --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": "SC117/Agents-A1-MTP-APEX-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "SC117/Agents-A1-MTP-APEX-GGUF" \
        --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": "SC117/Agents-A1-MTP-APEX-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use SC117/Agents-A1-MTP-APEX-GGUF with Ollama:
```
ollama run hf.co/SC117/Agents-A1-MTP-APEX-GGUF:BF16
```

Unsloth Studio

How to use SC117/Agents-A1-MTP-APEX-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 SC117/Agents-A1-MTP-APEX-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 SC117/Agents-A1-MTP-APEX-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for SC117/Agents-A1-MTP-APEX-GGUF to start chatting

How to use SC117/Agents-A1-MTP-APEX-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf SC117/Agents-A1-MTP-APEX-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": "SC117/Agents-A1-MTP-APEX-GGUF:BF16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use SC117/Agents-A1-MTP-APEX-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 SC117/Agents-A1-MTP-APEX-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 SC117/Agents-A1-MTP-APEX-GGUF:BF16

Run Hermes

hermes

Atomic Chat new

OpenClaw new

How to use SC117/Agents-A1-MTP-APEX-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf SC117/Agents-A1-MTP-APEX-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 "SC117/Agents-A1-MTP-APEX-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"

Docker Model Runner
How to use SC117/Agents-A1-MTP-APEX-GGUF with Docker Model Runner:
```
docker model run hf.co/SC117/Agents-A1-MTP-APEX-GGUF:BF16
```

Lemonade

How to use SC117/Agents-A1-MTP-APEX-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull SC117/Agents-A1-MTP-APEX-GGUF:BF16

Run and chat with the model

lemonade run user.Agents-A1-MTP-APEX-GGUF-BF16

List all available models

lemonade list

SC117 commited on about 12 hours ago

Commit

68b12b8

verified ·

1 Parent(s): 995a934

Upload 2 files

Browse files

Files changed (2) hide show

README.md +14 -14
README_zh.md +14 -14

README.md CHANGED Viewed

@@ -65,7 +65,7 @@ base_model:
 <div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;">
 <p style="margin: 0 0 12px 0;">The released <a href="https://huggingface.co/InternScience/Agents-A1" target="_blank" style="color: #047857; text-decoration: none; font-weight: 700;">InternScience/Agents-A1</a> checkpoint is a <b>40-layer Qwen3.5-35B-A3B MoE</b> without MTP (Multi-Token Prediction) layers. To enable MTP acceleration in llama.cpp (which speeds up long-context generation by 10–30%), we <b>extract the 1 MTP layer from Qwen3.5-35B-A3B</b> and inject it into Agents-A1's safetensors before GGUF conversion.</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">Step 1 — Extract MTP tensors from Qwen3.5-35B-A3B</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;"># Source: J:\Models\Qwen3.5-35B-A3B-MTP (Qwen3.5-35B-A3B + native MTP)
 from safetensors import safe_open
 import json, os
@@ -73,27 +73,27 @@ src = r"J:\Models\Qwen3.5-35B-A3B-MTP"
 with open(os.path.join(src, "model.safetensors.index.json")) as f:
     idx = json.load(f)
 mtp_keys = [k for k in idx["weight_map"] if "mtp" in k.lower()]
-print(f"Found {len(mtp_keys)} MTP tensors")  # 785</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">Step 2 — Add as a new safetensors shard (N+1)</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;"># Save 785 MTP tensors as a new shard
 new_shard = "model.safetensors-15-of-15.safetensors"
 save_file({k: get_tensor(k) for k in mtp_keys}, new_shard)
 # Update model.safetensors.index.json:
 #   - metadata.total_size += new_shard_size
 #   - weight_map: append new_shard path for each MTP key
-#   - DO NOT modify existing 14 shards (avoid touching original data)</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">Step 3 — Convert HF → BF16 GGUF with master llama.cpp</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;">F:\llama.cpp\llama.cpp-master\convert_hf_to_gguf.py ^
   J:\Models\Agents-A1 ^
   --outfile J:\Models\Agents-A1-MTP-GGUF\Agents-A1-MTP-BF16.gguf ^
   --outtype f16
 # Master version handles Qwen3.5MoE with MTP auto:
 #   - Normal layers: blk.0–39
-#   - MTP layer: blk.40.nextn.*  (785 tensors)</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">Step 4 — Quantize with APEX (Q4_K_M default, MTP at Q8_0)</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;">F:\llama.cpp\...\llama-quantize.exe ^
   --imatrix J:\Models\Qwen3.5-35B-A3B.imatrix.gguf ^
   --tensor-type-file E:\apex-quant\configs\qwen36_35b_mtp_&lt;tier&gt;.txt ^
   J:\Models\Agents-A1-MTP-GGUF\Agents-A1-MTP-BF16.gguf ^
@@ -101,7 +101,7 @@ save_file({k: get_tensor(k) for k in mtp_keys}, new_shard)
   Q4_K_M
 # APEX qwen36_35b_mtp_*.txt configs include blk.40 overrides
-# (Q8_0 for MTP across all tiers) — no manual patching needed.</p>
 </div>
 </div>
@@ -120,17 +120,17 @@ save_file({k: get_tensor(k) for k in mtp_keys}, new_shard)
 <div style="background: linear-gradient(135deg, #10b981 0%, #059669 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🚀</span> Usage</div>
 <div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;">
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">llama.cpp (text only)</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">hf download SC117/Agents-A1-MTP-APEX-GGUF --include "*.gguf" --local-dir ./models
-./llama-server -m ./models/Agents-A1-MTP-APEX-I-Compact.gguf -ngl 99 -c 131072</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">llama.cpp (vision + text)</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">./llama-server -m ./models/Agents-A1-MTP-APEX-I-Compact.gguf --mmproj ./models/mmproj-F16.gguf -ngl 99 -c 131072</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">vLLM</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">vllm serve SC117/Agents-A1-MTP-APEX-GGUF --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3
 # Tool-call variant
-vllm serve SC117/Agents-A1-MTP-APEX-GGUF --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">SGLang</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">python3 -m sglang.launch_server --model-path "SC117/Agents-A1-MTP-APEX-GGUF" --host 0.0.0.0 --port 30000</p>
 </div>
 </div>

 <div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;">
 <p style="margin: 0 0 12px 0;">The released <a href="https://huggingface.co/InternScience/Agents-A1" target="_blank" style="color: #047857; text-decoration: none; font-weight: 700;">InternScience/Agents-A1</a> checkpoint is a <b>40-layer Qwen3.5-35B-A3B MoE</b> without MTP (Multi-Token Prediction) layers. To enable MTP acceleration in llama.cpp (which speeds up long-context generation by 10–30%), we <b>extract the 1 MTP layer from Qwen3.5-35B-A3B</b> and inject it into Agents-A1's safetensors before GGUF conversion.</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">Step 1 — Extract MTP tensors from Qwen3.5-35B-A3B</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;"># Source: J:\Models\Qwen3.5-35B-A3B-MTP (Qwen3.5-35B-A3B + native MTP)
 from safetensors import safe_open
 import json, os
 with open(os.path.join(src, "model.safetensors.index.json")) as f:
     idx = json.load(f)
 mtp_keys = [k for k in idx["weight_map"] if "mtp" in k.lower()]
+print(f"Found {len(mtp_keys)} MTP tensors")  # 785</pre>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">Step 2 — Add as a new safetensors shard (N+1)</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;"># Save 785 MTP tensors as a new shard
 new_shard = "model.safetensors-15-of-15.safetensors"
 save_file({k: get_tensor(k) for k in mtp_keys}, new_shard)
 # Update model.safetensors.index.json:
 #   - metadata.total_size += new_shard_size
 #   - weight_map: append new_shard path for each MTP key
+#   - DO NOT modify existing 14 shards (avoid touching original data)</pre>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">Step 3 — Convert HF → BF16 GGUF with master llama.cpp</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;">F:\llama.cpp\llama.cpp-master\convert_hf_to_gguf.py ^
   J:\Models\Agents-A1 ^
   --outfile J:\Models\Agents-A1-MTP-GGUF\Agents-A1-MTP-BF16.gguf ^
   --outtype f16
 # Master version handles Qwen3.5MoE with MTP auto:
 #   - Normal layers: blk.0–39
+#   - MTP layer: blk.40.nextn.*  (785 tensors)</pre>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">Step 4 — Quantize with APEX (Q4_K_M default, MTP at Q8_0)</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;">F:\llama.cpp\...\llama-quantize.exe ^
   --imatrix J:\Models\Qwen3.5-35B-A3B.imatrix.gguf ^
   --tensor-type-file E:\apex-quant\configs\qwen36_35b_mtp_&lt;tier&gt;.txt ^
   J:\Models\Agents-A1-MTP-GGUF\Agents-A1-MTP-BF16.gguf ^
   Q4_K_M
 # APEX qwen36_35b_mtp_*.txt configs include blk.40 overrides
+# (Q8_0 for MTP across all tiers) — no manual patching needed.</pre>
 </div>
 </div>
 <div style="background: linear-gradient(135deg, #10b981 0%, #059669 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🚀</span> Usage</div>
 <div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;">
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">llama.cpp (text only)</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">hf download SC117/Agents-A1-MTP-APEX-GGUF --include "*.gguf" --local-dir ./models
+./llama-server -m ./models/Agents-A1-MTP-APEX-I-Compact.gguf -ngl 99 -c 131072</pre>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">llama.cpp (vision + text)</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">./llama-server -m ./models/Agents-A1-MTP-APEX-I-Compact.gguf --mmproj ./models/mmproj-F16.gguf -ngl 99 -c 131072</pre>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">vLLM</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">vllm serve SC117/Agents-A1-MTP-APEX-GGUF --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3
 # Tool-call variant
+vllm serve SC117/Agents-A1-MTP-APEX-GGUF --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder</pre>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">SGLang</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">python3 -m sglang.launch_server --model-path "SC117/Agents-A1-MTP-APEX-GGUF" --host 0.0.0.0 --port 30000</pre>
 </div>
 </div>

README_zh.md CHANGED Viewed

@@ -65,7 +65,7 @@ base_model:
 <div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;">
 <p style="margin: 0 0 12px 0;">官方发布的 <a href="https://huggingface.co/InternScience/Agents-A1" target="_blank" style="color: #047857; text-decoration: none; font-weight: 700;">InternScience/Agents-A1</a> checkpoint 是一个 <b>40 层 Qwen3.5-35B-A3B MoE</b>，不包含 MTP（Multi-Token Prediction）层。为了在 llama.cpp 中启用 MTP 加速（长上下文生成提速 10–30%），我们 <b>从 Qwen3.5-35B-A3B 中提取 1 层 MTP</b>，注入到 Agents-A1 的 safetensors 中，再转 GGUF。</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">步骤 1 — 从 Qwen3.5-35B-A3B 提取 MTP tensor</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;"># 源：J:\Models\Qwen3.5-35B-A3B-MTP（Qwen3.5-35B-A3B + 原生 MTP）
 from safetensors import safe_open
 import json, os
@@ -73,27 +73,27 @@ src = r"J:\Models\Qwen3.5-35B-A3B-MTP"
 with open(os.path.join(src, "model.safetensors.index.json")) as f:
     idx = json.load(f)
 mtp_keys = [k for k in idx["weight_map"] if "mtp" in k.lower()]
-print(f"Found {len(mtp_keys)} MTP tensors")  # 785</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">步骤 2 — 作为新分片（N+1）追加</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;"># 把 785 个 MTP tensor 保存为新分片
 new_shard = "model.safetensors-15-of-15.safetensors"
 save_file({k: get_tensor(k) for k in mtp_keys}, new_shard)
 # 更新 model.safetensors.index.json：
 #   - metadata.total_size += 新分片大小
 #   - weight_map: 为每个 MTP key 追加新分片路径
-#   - 不修改原 14 个分片（避免触碰原始数据）</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">步骤 3 — 用 master llama.cpp 转 BF16 GGUF</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;">F:\llama.cpp\llama.cpp-master\convert_hf_to_gguf.py ^
   J:\Models\Agents-A1 ^
   --outfile J:\Models\Agents-A1-MTP-GGUF\Agents-A1-MTP-BF16.gguf ^
   --outtype f16
 # master 版本自动处理 Qwen3.5MoE + MTP：
 #   - 常规层：blk.0–39
-#   - MTP 层：blk.40.nextn.*  （785 个 tensor）</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">步骤 4 — 用 APEX 量化（Q4_K_M 默认，MTP 用 Q8_0）</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;">F:\llama.cpp\...\llama-quantize.exe ^
   --imatrix J:\Models\Qwen3.5-35B-A3B.imatrix.gguf ^
   --tensor-type-file E:\apex-quant\configs\qwen36_35b_mtp_&lt;档位&gt;.txt ^
   J:\Models\Agents-A1-MTP-GGUF\Agents-A1-MTP-BF16.gguf ^
@@ -101,7 +101,7 @@ save_file({k: get_tensor(k) for k in mtp_keys}, new_shard)
   Q4_K_M
 # APEX qwen36_35b_mtp_*.txt 配置已包含 blk.40 override
-# （所有档位 MTP 用 Q8_0）—— 无需手动 patch。</p>
 </div>
 </div>
@@ -120,17 +120,17 @@ save_file({k: get_tensor(k) for k in mtp_keys}, new_shard)
 <div style="background: linear-gradient(135deg, #10b981 0%, #059669 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🚀</span> 使用方法</div>
 <div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;">
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">llama.cpp（纯文本）</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">hf download SC117/Agents-A1-MTP-APEX-GGUF --include "*.gguf" --local-dir ./models
-./llama-server -m ./models/Agents-A1-MTP-APEX-I-Compact.gguf -ngl 99 -c 131072</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">llama.cpp（视觉 + 文本）</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">./llama-server -m ./models/Agents-A1-MTP-APEX-I-Compact.gguf --mmproj ./models/mmproj-F16.gguf -ngl 99 -c 131072</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">vLLM</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">vllm serve SC117/Agents-A1-MTP-APEX-GGUF --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3
 # 工具调用变体
-vllm serve SC117/Agents-A1-MTP-APEX-GGUF --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">SGLang</p>
-<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">python3 -m sglang.launch_server --model-path "SC117/Agents-A1-MTP-APEX-GGUF" --host 0.0.0.0 --port 30000</p>
 </div>
 </div>

 <div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;">
 <p style="margin: 0 0 12px 0;">官方发布的 <a href="https://huggingface.co/InternScience/Agents-A1" target="_blank" style="color: #047857; text-decoration: none; font-weight: 700;">InternScience/Agents-A1</a> checkpoint 是一个 <b>40 层 Qwen3.5-35B-A3B MoE</b>，不包含 MTP（Multi-Token Prediction）层。为了在 llama.cpp 中启用 MTP 加速（长上下文生成提速 10–30%），我们 <b>从 Qwen3.5-35B-A3B 中提取 1 层 MTP</b>，注入到 Agents-A1 的 safetensors 中，再转 GGUF。</p>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">步骤 1 — 从 Qwen3.5-35B-A3B 提取 MTP tensor</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;"># 源：J:\Models\Qwen3.5-35B-A3B-MTP（Qwen3.5-35B-A3B + 原生 MTP）
 from safetensors import safe_open
 import json, os
 with open(os.path.join(src, "model.safetensors.index.json")) as f:
     idx = json.load(f)
 mtp_keys = [k for k in idx["weight_map"] if "mtp" in k.lower()]
+print(f"Found {len(mtp_keys)} MTP tensors")  # 785</pre>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">步骤 2 — 作为新分片（N+1）追加</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;"># 把 785 个 MTP tensor 保存为新分片
 new_shard = "model.safetensors-15-of-15.safetensors"
 save_file({k: get_tensor(k) for k in mtp_keys}, new_shard)
 # 更新 model.safetensors.index.json：
 #   - metadata.total_size += 新分片大小
 #   - weight_map: 为每个 MTP key 追加新分片路径
+#   - 不修改原 14 个分片（避免触碰原始数据）</pre>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">步骤 3 — 用 master llama.cpp 转 BF16 GGUF</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;">F:\llama.cpp\llama.cpp-master\convert_hf_to_gguf.py ^
   J:\Models\Agents-A1 ^
   --outfile J:\Models\Agents-A1-MTP-GGUF\Agents-A1-MTP-BF16.gguf ^
   --outtype f16
 # master 版本自动处理 Qwen3.5MoE + MTP：
 #   - 常规层：blk.0–39
+#   - MTP 层：blk.40.nextn.*  （785 个 tensor）</pre>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">步骤 4 — 用 APEX 量化（Q4_K_M 默认，MTP 用 Q8_0）</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre;">F:\llama.cpp\...\llama-quantize.exe ^
   --imatrix J:\Models\Qwen3.5-35B-A3B.imatrix.gguf ^
   --tensor-type-file E:\apex-quant\configs\qwen36_35b_mtp_&lt;档位&gt;.txt ^
   J:\Models\Agents-A1-MTP-GGUF\Agents-A1-MTP-BF16.gguf ^
   Q4_K_M
 # APEX qwen36_35b_mtp_*.txt 配置已包含 blk.40 override
+# （所有档位 MTP 用 Q8_0）—— 无需手动 patch。</pre>
 </div>
 </div>
 <div style="background: linear-gradient(135deg, #10b981 0%, #059669 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🚀</span> 使用方法</div>
 <div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;">
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">llama.cpp（纯文本）</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">hf download SC117/Agents-A1-MTP-APEX-GGUF --include "*.gguf" --local-dir ./models
+./llama-server -m ./models/Agents-A1-MTP-APEX-I-Compact.gguf -ngl 99 -c 131072</pre>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">llama.cpp（视觉 + 文本）</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">./llama-server -m ./models/Agents-A1-MTP-APEX-I-Compact.gguf --mmproj ./models/mmproj-F16.gguf -ngl 99 -c 131072</pre>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">vLLM</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">vllm serve SC117/Agents-A1-MTP-APEX-GGUF --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3
 # 工具调用变体
+vllm serve SC117/Agents-A1-MTP-APEX-GGUF --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder</pre>
 <p style="margin: 0 0 8px 0; font-weight: bold; color: #064e3b;">SGLang</p>
+<pre style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">python3 -m sglang.launch_server --model-path "SC117/Agents-A1-MTP-APEX-GGUF" --host 0.0.0.0 --port 30000</pre>
 </div>
 </div>