Fix export models for image generation (#3508) (#3514)

Fixing issue introduced with: d981ff0. When performing quantization after switching from int8 to int4 in export_models.py we encountered issue:
"Channel size 64 should be divisible by size of group 128."

Cherry-pick of:c55551763d02825829337b62c2dcef9339706f79

Ticket:CVS-169363

Author: atobiszei

demos/common/export_models/export_model.py

@@ -30,6 +30,7 @@ def add_common_arguments(parser):
     parser.add_argument('--overwrite_models', default=False, action='store_true', help='Overwrite the model if it already exists in the models repository', dest='overwrite_models')
     parser.add_argument('--target_device', default="CPU", help='CPU, GPU, NPU or HETERO, default is CPU', dest='target_device')
     parser.add_argument('--ov_cache_dir', default=None, help='Folder path for compilation cache to speedup initialization time', dest='ov_cache_dir')
+    parser.add_argument('--extra_quantization_params', required=False, help='Add advanced quantization parameters. Check optimum-intel documentation. Example: "--sym --group-size -1 --ratio 1.0 --awq --scale-estimation --dataset wikitext2"', dest='extra_quantization_params')
 
 parser = argparse.ArgumentParser(description='Export Hugging face models to OVMS models repository including all configuration for deployments')
 
@@ -38,7 +39,6 @@ def add_common_arguments(parser):
 add_common_arguments(parser_text)
 parser_text.add_argument('--pipeline_type', default=None, choices=["LM", "LM_CB", "VLM", "VLM_CB", "AUTO"], help='Type of the pipeline to be used. AUTO is used by default', dest='pipeline_type')
 parser_text.add_argument('--kv_cache_precision', default=None, choices=["u8"], help='u8 or empty (model default). Reduced kv cache precision to u8 lowers the cache size consumption.', dest='kv_cache_precision')
-parser_text.add_argument('--extra_quantization_params', help='Add advanced quantization parameters. Check optimum-intel documentation. Example: "--sym --group-size -1 --ratio 1.0 --awq --scale-estimation --dataset wikitext2"', dest='extra_quantization_params')
 parser_text.add_argument('--enable_prefix_caching', action='store_true', help='This algorithm is used to cache the prompt tokens.', dest='enable_prefix_caching')
 parser_text.add_argument('--disable_dynamic_split_fuse', action='store_false', help='The maximum number of tokens that can be batched together.', dest='dynamic_split_fuse')
 parser_text.add_argument('--max_num_batched_tokens', default=None, help='empty or integer. The maximum number of tokens that can be batched together.', dest='max_num_batched_tokens')
@@ -379,11 +379,9 @@ def export_text_generation_model(model_repository_path, source_model, model_name
                 if precision != 'int4':
                     print("NPU target device requires int4 precision. Changing to int4")
                     precision = 'int4'
-                if task_parameters['extra_quantization_params'] is None:
+                if task_parameters['extra_quantization_params'] == "":
                     print("Using default quantization parameters for NPU: --sym --ratio 1.0 --group-size -1")
                     task_parameters['extra_quantization_params'] = "--sym --ratio 1.0 --group-size -1"
-            if task_parameters['extra_quantization_params'] is None:
-                task_parameters['extra_quantization_params'] = ""
             optimum_command = "optimum-cli export openvino --model {} --weight-format {} {} --trust-remote-code {}".format(source_model, precision, task_parameters['extra_quantization_params'], llm_model_path)
             if os.system(optimum_command):
                 raise ValueError("Failed to export llm model", source_model)    
@@ -478,7 +476,7 @@ def export_embeddings_model(model_repository_path, source_model, model_name, pre
             embeddings_path = os.path.join(model_repository_path, model_name,'embeddings', version)
             print("Exporting embeddings model to ",embeddings_path)
             if not os.path.isdir(embeddings_path) or args['overwrite_models']:
-                optimum_command = "optimum-cli export openvino --disable-convert-tokenizer --model {} --task feature-extraction --weight-format {} --trust-remote-code --library sentence_transformers {}".format(source_model, precision, tmpdirname)
+                optimum_command = "optimum-cli export openvino --disable-convert-tokenizer --model {} --task feature-extraction --weight-format {} {} --trust-remote-code --library sentence_transformers {}".format(source_model, precision, task_parameters['extra_quantization_params'], tmpdirname)
                 if os.system(optimum_command):
                     raise ValueError("Failed to export embeddings model", source_model)
                 set_rt_info(tmpdirname, 'openvino_model.xml', 'config.json')
@@ -517,7 +515,7 @@ def export_embeddings_model_ov(model_repository_path, source_model, model_name,
     destination_path = os.path.join(model_repository_path, model_name)
     print("Exporting embeddings model to ",destination_path)
     if not os.path.isdir(destination_path) or args['overwrite_models']:
-        optimum_command = "optimum-cli export openvino --model {} --disable-convert-tokenizer --task feature-extraction --weight-format {} --trust-remote-code --library sentence_transformers {}".format(source_model, precision, destination_path)
+        optimum_command = "optimum-cli export openvino --model {} --disable-convert-tokenizer --task feature-extraction --weight-format {} {} --trust-remote-code --library sentence_transformers {}".format(source_model, precision, task_parameters['extra_quantization_params'], destination_path)
         if os.system(optimum_command):
             raise ValueError("Failed to export embeddings model", source_model)
         if truncate:
@@ -539,7 +537,7 @@ def export_rerank_model_ov(model_repository_path, source_model, model_name, prec
     destination_path = os.path.join(model_repository_path, model_name)
     print("Exporting rerank model to ",destination_path)
     if not os.path.isdir(destination_path) or args['overwrite_models']:
-        optimum_command = "optimum-cli export openvino --model {} --disable-convert-tokenizer --task text-classification --weight-format {} --trust-remote-code {}".format(source_model, precision, destination_path)
+        optimum_command = "optimum-cli export openvino --model {} --disable-convert-tokenizer --task text-classification --weight-format {} {} --trust-remote-code {}".format(source_model, precision, task_parameters['extra_quantization_params'], destination_path)
         if os.system(optimum_command):
             raise ValueError("Failed to export rerank model", source_model)
         print("Exporting tokenizer to ", destination_path)
@@ -565,7 +563,7 @@ def export_rerank_model(model_repository_path, source_model, model_name, precisi
             embeddings_path = os.path.join(model_repository_path, model_name, 'rerank', version)
             print("Exporting rerank model to ",embeddings_path)
             if not os.path.isdir(embeddings_path) or args['overwrite_models']:
-                optimum_command = "optimum-cli export openvino --disable-convert-tokenizer --model {} --task text-classification --weight-format {} --trust-remote-code {}".format(source_model, precision, tmpdirname)
+                optimum_command = "optimum-cli export openvino --disable-convert-tokenizer --model {} --task text-classification --weight-format {} {} --trust-remote-code {}".format(source_model, precision, task_parameters['extra_quantization_params'], tmpdirname)
                 if os.system(optimum_command):
                     raise ValueError("Failed to export rerank model", source_model)
                 set_rt_info(tmpdirname, 'openvino_model.xml', 'config.json')
@@ -601,7 +599,8 @@ def export_image_generation_model(model_repository_path, source_model, model_nam
     if os.path.isfile(model_index_path):
         print("Model index file already exists. Skipping conversion, re-generating graph only.")
     else:
-        optimum_command = "optimum-cli export openvino --model {} --weight-format {} {}".format(source_model, precision, target_path)
+        optimum_command = "optimum-cli export openvino --model {} --weight-format {} {} {}".format(source_model, precision, task_parameters['extra_quantization_params'], target_path)
+        print(f'optimum cli command: {optimum_command}')
         if os.system(optimum_command):
             raise ValueError("Failed to export image generation model", source_model)
 
@@ -653,6 +652,8 @@ def export_image_generation_model(model_repository_path, source_model, model_nam
 template_parameters = {k: v for k, v in args.items() if k not in ['model_repository_path', 'source_model', 'model_name', 'precision', 'version', 'config_file_path', 'overwrite_models']}
 print("template params:", template_parameters)
 
+if template_parameters['extra_quantization_params'] is None:
+    template_parameters['extra_quantization_params'] = ""
 if args['task'] == 'text_generation':
     export_text_generation_model(args['model_repository_path'], args['source_model'], args['model_name'], args['precision'], template_parameters, args['config_file_path'])
 
@@ -677,5 +678,6 @@ def export_image_generation_model(model_repository_path, source_model, model_nam
         'max_num_images_per_prompt',
         'default_num_inference_steps',
         'max_num_inference_steps',
+        'extra_quantization_params'
     ]}
     export_image_generation_model(args['model_repository_path'], args['source_model'], args['model_name'], args['precision'], template_parameters, args['config_file_path'], args['num_streams'])

demos/image_generation/README.md

@@ -127,13 +127,16 @@ Run `export_model.py` script to download and quantize the model:
 
 > **Note:** The users in China need to set environment variable HF_ENDPOINT="https://hf-mirror.com" before running the export script to connect to the HF Hub.
 
+> **Note:** The `--extra_quantization_params` parameter is used to pass additional parameters to the optimum-cli. It may be required to set the `--group-size` parameter when quantizing the model when encountering errors like: `Channel size 64 should be divisible by size of group 128.`
+
 ### Export model for CPU
 ```console
 python export_model.py image_generation \
   --source_model black-forest-labs/FLUX.1-schnell \
   --weight-format int4 \
   --config_file_path models/config.json \
   --model_repository_path models \
+  --extra_quantization_params "--group-size 64" \
   --overwrite_models
 ```
 
@@ -145,6 +148,7 @@ python export_model.py image_generation \
   --target_device GPU \
   --config_file_path models/config.json \
   --model_repository_path models \
+  --extra_quantization_params "--group-size 64" \
   --overwrite_models
 ```
 
@@ -209,7 +213,7 @@ In case you want to use GPU device to run the generation, add extra docker param
 to `docker run` command, use the image with GPU support. Export the models with precision matching the GPU capacity and adjust pipeline configuration.
 It can be applied using the commands below:
 ```bash
-docker run -d --rm -p 8000:8000 -v $(pwd)/models:/workspace:ro \
+docker run -d --rm -p 8000:8000 -v $(pwd)/models:/models:ro \
   --device /dev/dri --group-add=$(stat -c "%g" /dev/dri/render* | head -n 1) \
   openvino/model_server:2025.2-gpu \
     --rest_port 8000 \
@@ -233,7 +237,6 @@ ovms --rest_port 8000 ^
 
 ::::
 
-
 ## Readiness Check
 
 Wait for the model to load. You can check the status with a simple command: