[ Don't Pull ] - Adds full support for converting Qwen3.6 35B A3B MoE variants

.gitignore

@@ -1,6 +1,7 @@
 # Byte-compiled / optimized / DLL files
 gguf_files/
 q4nx_files/
+log/
 *.ipynb
 *.bin
 *.gguf

README.md

@@ -1,6 +1,6 @@
 # FLM Q4NX Converter
 
-A utility for converting GGUF model files into the Q4NX format. This tool supports converting language, vision, and audio model weights.
+A utility for converting GGUF model files, usually fine-tunes, into the Q4NX format. This tool supports converting language, vision, and audio model weights.
 
 ## Supported Models
 Based on the configuration, the converter supports several model architectures, including:
@@ -14,25 +14,26 @@ Based on the configuration, the converter supports several model architectures,
 - Qwen 3
 - Qwen 3 VL
 - Qwen 3.5
+- Qwen 3.5 MoE
+- Qwen 3.6 MoE
 
 ### Weight Type Support
 - `language`: supported for model families in `configs/`
 - `vision`: supported for vision-capable architectures (for example, Gemma 4 and Qwen3-VL)
-- `audio`: currently supported for **Gemma 4** (`-t audio`)
+- `audio`: currently supported for **Gemma 4** variants (`-t audio`)
 
 ## Setup
 
-The project includes a setup script that automatically creates a Python virtual environment, installs required dependencies (like `torch`, `amd-quark`, `transformers`), and sets up the `gguf` package from `llama.cpp`.
+Create a virtual environment using python 3.13
 
-1. Make the setup script executable (if it isn't already) and run it:
+1. Create and activate a virtual environment using pip or uv:
    ```bash
-   chmod +x setup_venv.sh
-   ./setup_venv.sh
+   uv venv venv
+   venv/bin/activate
    ```
-
-2. Activate the environment using the provided activation script. **Note:** Always use `activate.sh` instead of directly sourcing the venv, as it also sets up necessary environment variables like `HF_HOME` and CUDA paths.
+2. Install dependencies (one time)
    ```bash
-   source activate.sh
+   uv pip install -r requirements.txt
    ```
 
 ## Usage

configs/qwen2vl.json

@@ -7,7 +7,8 @@
     },
     "vision_config":{
         "vision_MM_K": 256,
-        "vision_MM_N": 64
+        "vision_MM_N": 64,
+        "vision_file": "vision_weights.q4nx"
     },        
     "default_tensor_type": "Q4_1",
     "name_map": {

configs/qwen35moe.json

@@ -0,0 +1,15 @@
+{
+  "q4nx_config": {
+    "row_block_size": 32,
+    "col_block_size": 256,
+    "parallel_size": 16,
+    "keep_block_in_2D": true
+  },
+  "default_tensor_type": "Q4_1",
+  "name_map": {},
+  "addr_qk": 9216,
+  "addr_kv": 36864,
+  "addr_kk": 45056,
+  "addr_l_begin_mha": 53760,
+  "addr_l_end_mha": 25088
+}

convert.py

@@ -1,35 +1,138 @@
-#!./venv/bin/python3
+#!/usr/bin/env python3
 import argparse
 import os
 from pathlib import Path
-from q4nx import create_converter
+from q4nx import create_converter, create_hf_converter
+from q4nx.model_assets import assemble_model_assets, assemble_model_assets_hf, get_default_flm_version, find_repo_gguf
 
 
-def convert_gguf_to_q4nx(gguf_path: str, q4nx_path: str, override_model_arch:str, weights_type: str = 'language'):
-    model = create_converter(gguf_path, override_model_arch)
-    model.convert(q4nx_path=q4nx_path, weights_type=weights_type)
+def is_hf_repo_id(path: str) -> bool:
+    """HF hub repo id like 'org/name' (not a local path, not a .gguf)."""
+    if not path or path.endswith(".gguf") or os.path.exists(path):
+        return False
+    if path.startswith(("http://", "https://", "file:")):
+        return False
+    # org/name with no filesystem separators beyond the single slash
+    parts = path.split("/")
+    return len(parts) == 2 and all(parts) and "\\" not in path
+
+
+def is_hf_repo_id(path: str) -> bool:
+    """HF hub repo id like 'org/name' (not a local path, not a .gguf)."""
+    if not path or path.endswith(".gguf") or os.path.exists(path):
+        return False
+    if path.startswith(("http://", "https://", "file:")):
+        return False
+    # org/name with no filesystem separators beyond the single slash
+    parts = path.split("/")
+    return len(parts) == 2 and all(parts) and "\\" not in path
+
+
+def is_hf_source(path: str) -> bool:
+    if os.path.isdir(path):
+        return (
+            os.path.exists(os.path.join(path, "model.safetensors"))
+            or os.path.exists(os.path.join(path, "model.safetensors.index.json"))
+        )
+    return is_hf_repo_id(path)
+
+
+def convert_gguf_to_q4nx(gguf_path: str, q4nx_path: str, override_model_arch:str, weights_type: str = 'language', source_model: str = None, flm_version: str = None, deploy_tag: str = None, deploy_from: str = None, deploy_name: str = None):
+    if flm_version is None:
+        flm_version = get_default_flm_version()
+    # The weight source is always -i: an HF dir / repo id (Darwin-style HF
+    # path) or a GGUF. --source-model only supplies tokenizer/config assets and
+    # is never treated as a weight source, so it can't trigger a weights
+    # download when converting from GGUF.
+    #
+    # New: -i <hf-repo-id> prefers a quantized GGUF shipped in the repo itself,
+    # chosen in a family-preferred order (default q4_1, then q4_0, then q8_0;
+    # e.g. LFM prefers q4_0 first, gpt-oss also accepts mxfp4 last). The order
+    # is driven by -f when given, otherwise by a best-effort repo-id/filename
+    # match. The chosen GGUF is downloaded via the HF cache; if the repo has
+    # none, we fall back to the HF-safetensors source path below.
+    hf_input = None
+    source_file = None
+    if is_hf_repo_id(gguf_path):
+        repo_id = gguf_path
+        found = find_repo_gguf(repo_id, override_model_arch)
+        if found is not None:
+            gguf_path, source_file = found
+            source_model = source_model or repo_id
+        else:
+            hf_input = repo_id
+    elif is_hf_source(gguf_path):
+        hf_input = gguf_path
+
+    if hf_input is not None:
+        model = create_hf_converter(hf_input, override_model_arch)
+        model.convert(q4nx_path=q4nx_path, weights_type=weights_type)
+        assemble_model_assets_hf(
+            model.hf_source,
+            model.q4nx_config,
+            q4nx_path,
+            source_model=source_model or hf_input,
+            flm_version=flm_version,
+            source_file=source_file,
+        )
+    else:
+        model = create_converter(gguf_path, override_model_arch)
+        model.convert(q4nx_path=q4nx_path, weights_type=weights_type)
+        assemble_model_assets(
+            model.gguf_reader,
+            model.q4nx_config,
+            q4nx_path,
+            source_model=source_model,
+            flm_version=flm_version,
+            source_file=source_file,
+        )
+    if deploy_tag:
+        from q4nx.deploy import deploy_model
+        deploy_model(
+            q4nx_path,
+            deploy_tag,
+            model.model_arch,
+            model_dir_name=deploy_name,
+            deploy_from=deploy_from,
+        )
+    return model
 
 
 def main():
     parser = argparse.ArgumentParser(
-        description='Convert GGUF model files to Q4NX format (output always named model.q4nx)',
+        description='Convert GGUF or HF-safetensors model files to Q4NX format (output always named model.q4nx). '
+                    '-i also accepts an HF repo id: the repo is searched for a quantized GGUF, chosen in a '
+                    'family-preferred order (default q4_1 / q4_0 / q8_0; e.g. LFM prefers q4_0 first). Pass -f to '
+                    'force the family and override the auto-detected source GGUF.',
         epilog='Examples:\n'
                '  python convert.py -i model.gguf\n'
                '  python convert.py -i model.gguf -o output_folder\n'
                '  python convert.py model.gguf output_folder\n'
-               '  python convert.py model.gguf .\n'
+               '  python convert.py -i Qwen/Qwen3.5-9B -o output_folder     (HF repo: picks a q4_1/q4_0/q8_0 GGUF from it)\n'
+               '  python convert.py -i LiquidAI/LFM2-1.2B -o out -f lfm2      (force LFM family -> prefers q4_0 source)\n'
+               '  python convert.py -i /path/to/hf_model_dir -o output_folder   (HF safetensors source, e.g. Darwin-36B-Opus)\n'
                '  python convert.py -i vision_model.gguf -o output_folder -t vision',
         formatter_class=argparse.RawDescriptionHelpFormatter
     )
 
     # Add support for both flag-based and positional arguments
     parser.add_argument('input_file', nargs='?', help='Input GGUF file (positional)')
     parser.add_argument('output_folder', nargs='?', help='Output folder (positional, optional)')
-    parser.add_argument('-i', '--input', dest='input_flag', help='Input GGUF file')
+    parser.add_argument('-i', '--input', dest='input_flag', help='Input GGUF file, or an HF repo id (a quantized GGUF is auto-selected in family-preferred order; use -f to force the family)')
     parser.add_argument('-o', '--output', dest='output_flag', help='Output folder (optional, defaults to input file directory)')
     parser.add_argument('-t', '--type', dest='weights_type', default='language', help='Type of weights to convert (default: language)',
                         choices=['language', 'vision', 'audio'])
     parser.add_argument('-f', '--force', dest='force_model_type', default="", help="Model type. Empty string for automatic recognition from gguf file")
+    parser.add_argument('-s', '--source-model', dest='source_model', default=None,
+                        help="Source HF/ModelScope model for tokenizer/config assets. A local dir, an HF cache repo name, or a repo id like 'Qwen/Qwen3.5-9B'. If omitted, the GGUF's provenance metadata is followed (local HF cache first, then SDK download).")
+    parser.add_argument('--flm-version', dest='flm_version', default=None,
+                        help="flm_version to write into the generated config.json (default: detected from `flm --version`)" )
+    parser.add_argument('-d', '--deploy', dest='deploy_tag', default=None, metavar='NAME:SIZE',
+                        help="Deploy the converted model into flm's models directory and register it under this tag (e.g. 'qwen3.5-claude:9b'). Uses a user-level model_list.json (point FLM_CONFIG_PATH at it to make `flm run` see the tag).")
+    parser.add_argument('--deploy-from', dest='deploy_from', default=None, metavar='SOURCE_TAG',
+                        help="Official registry entry to copy defaults from (e.g. 'qwen3.5:9b'). Auto-detected from the model architecture if omitted.")
+    parser.add_argument('--deploy-name', dest='deploy_name', default=None, metavar='DIR',
+                        help="Directory name inside flm's models dir (default: derived from the deploy tag, e.g. Qwen3.5-Claude-9B-NPU2).")
 
     args = parser.parse_args()
 
@@ -42,69 +145,22 @@ def main():
     # Determine output folder (prioritize flag, then positional)
     output_folder = args.output_flag or args.output_folder
 
-    # Check if input file exists
-    if not os.path.exists(input_path):
+    # Local paths must exist; HF repo ids are resolved later by the converter.
+    if not is_hf_repo_id(input_path) and not os.path.exists(input_path):
         parser.error(f'Input file does not exist: {input_path}')
-    
+
     # Create output directory if it doesn't exist
     output_dir = os.path.dirname(output_folder)
     if output_dir and not os.path.exists(output_dir):
         os.makedirs(output_dir, exist_ok=True)
 
     print(f"[INFO] Converting {input_path} to {output_folder}...")
-    convert_gguf_to_q4nx(input_path, output_folder, args.force_model_type, weights_type=args.weights_type)
+    convert_gguf_to_q4nx(input_path, output_folder, args.force_model_type, weights_type=args.weights_type, source_model=args.source_model, flm_version=args.flm_version, deploy_tag=args.deploy_tag, deploy_from=args.deploy_from, deploy_name=args.deploy_name)
     print(f"[INFO] Conversion complete! Output saved to {output_folder}")
 
 
 
 if __name__ == "__main__":
     # for debug, give the path and ouptut path here by directly set the command line args
     import sys
-    # sys.argv = ['convert.py', '-i', 'unsloth_gpt-oss-20b-Q4_0.gguf', '-o', 'unsloth-gotoss20b-q40']
-
-    # sys.argv = ['convert.py', '-i', 'unsloth_gpt-oss-20b-Q4_1.gguf', '-o', 'unsloth-gotoss20b-q41']            
-
-    # import sys
-    # # sys.argv = ['convert.py', '-i', 'gemma-3-4b-it-Q4_1.gguf', '-o', 'unsloth-gemma3-q41']    
-    # # main()
-
-
-    # # sys.argv = ['convert.py', '-i', 'gemma3-mmproj-BF16.gguf', '-o', 'unsloth-gemma3-vision', '-t', 'vision']
-    # # main()
-
-
-
-    # # sys.argv = ['convert.py', '-i'
-    # , 'medgemma3-mmproj-BF16.gguf', '-o', 'unsloth-medgemma3-vision', '-t', 'vision']
-    # # main()
-    # sys.argv = ['convert.py', '-i', 'Qwen3-VL-4B-Instruct-Q4_1.gguf', '-o', 'unsloth-qwen3vl-4b-q41' ]
-    # main()             
-
-    # sys.argv = ['convert.py', '-i', 'Qwen3-4B-Q4_1.gguf', '-o', 'unsloth-qwen3-4b-q41' ]
-    # main()                 
-    # sys.argv = ['convert.py', '-i', 'qwen3vl-4b-mmproj-BF16.gguf', '-o', 'unsloth-qwen3vl-vision', '-t', 'vision']
-    # main()        
-
-
-
-    #sys.argv = ['convert.py', '-i', 'qwen3_5vl-4b-mmproj-BF16.gguf', '-o', 'unsloth-qwen3_5vl-vision', '-t', 'vision']
-
-    #sys.argv = ['convert.py', '-i', 'qwen3_5vl-9bmmproj-BF16.gguf', '-o', 'unsloth-qwen3_5_9bvl-vision', '-t', 'vision'] 
-
-
-    #sys.argv = ['convert.py', '-i', 'Qwen3.5-4B-Q4_1.gguf', '-o', 'unsloth-qwen3_5_4bq41'] 
-
-    #sys.argv = ['convert.py', '-i', 'Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf', '-o', 'unsloth-qwen3_59b_uncensored', "-f", "qwen3.5-9B"]     
-    # sys.argv = ['convert.py', '-i', 'Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf', '-o', 'unsloth-qwen3_59b_uncensored', "-f", "qwen3.5-9B"]     
-
-    # sys.argv = ['convert.py', '-i', 'Qwen3.5-9B-Q4_1.gguf', '-o', 'unsloth-qwen3_5_9bq41']     
-
-
-
-    #sys.argv = ['convert.py', '-i', 'gemma-4-E2B-it-Q4_1.gguf', '-o', 'unsloth-gemma4-2b-it-q41']    
-
-    #sys.argv = ['convert.py', '-i', 'gemma4-2b-mmproj.gguf', '-o', 'unsloth-gemma4-2b-vision', '-t', 'vision']     
-
-    sys.argv = ['convert.py', '-i', 'gemma4-2b-mmproj.gguf', '-o', 'unsloth-gemma4-2b-audio', '-t', 'audio']         
-    # sys.argv = ['convert.py', '-i', 'debug_gemma4e2b_model.gguf', '-o', 'debug-gemma4-2b-audio', '-t', 'audio', '-f', 'gemma4']           
     main()