diff --git a/python/sglang/multimodal_gen/runtime/loader/component_loader.py b/python/sglang/multimodal_gen/runtime/loader/component_loader.py index 5e2175be3..9ad2b2753 100644 --- a/python/sglang/multimodal_gen/runtime/loader/component_loader.py +++ b/python/sglang/multimodal_gen/runtime/loader/component_loader.py @@ -23,6 +23,9 @@ from transformers import AutoImageProcessor, AutoProcessor, AutoTokenizer from transformers.utils import SAFE_WEIGHTS_INDEX_NAME from sglang.multimodal_gen.configs.models import EncoderConfig, ModelConfig +from sglang.multimodal_gen.configs.pipeline_configs.qwen_image import ( + QwenImageEditPipelineConfig, +) from sglang.multimodal_gen.runtime.distributed import get_local_torch_device from sglang.multimodal_gen.runtime.loader.fsdp_load import ( maybe_load_fsdp_model, @@ -462,6 +465,14 @@ class TextEncoderLoader(ComponentLoader): with local_torch_device, skip_init_modules(): architectures = getattr(model_config, "architectures", []) model_cls, _ = ModelRegistry.resolve_model_cls(architectures) + enable_image_understanding = ( + True + if isinstance( + server_args.pipeline_config, QwenImageEditPipelineConfig + ) + else False + ) + model_config.enable_image_understanding = enable_image_understanding model = model_cls(model_config) weights_to_load = {name for name, _ in model.named_parameters()} diff --git a/python/sglang/multimodal_gen/runtime/models/encoders/qwen2_5vl.py b/python/sglang/multimodal_gen/runtime/models/encoders/qwen2_5vl.py index c354f9237..364b72d59 100644 --- a/python/sglang/multimodal_gen/runtime/models/encoders/qwen2_5vl.py +++ b/python/sglang/multimodal_gen/runtime/models/encoders/qwen2_5vl.py @@ -1,7 +1,5 @@ # Copied and adapted from: https://github.com/hao-ai-lab/FastVideo -from types import SimpleNamespace - from transformers import ( Cache, DynamicCache, @@ -508,13 +506,15 @@ class Qwen2_5_VLModel(nn.Module): accepts_loss_kwargs = False _no_split_modules = ["Qwen2_5_VLDecoderLayer", "Qwen2_5_VLVisionBlock"] - def __init__(self, config): + def __init__(self, config, enable_image_understanding: bool = False): super().__init__() - self.visual = Qwen2_5_VisionTransformerPretrainedModel._from_config( - config.vision_config - ) self.language_model = Qwen2_5_VLTextModel(config.text_config) - self.visual.to(torch.get_default_dtype()) + + if enable_image_understanding: + self.visual = Qwen2_5_VisionTransformerPretrainedModel._from_config( + config.vision_config + ) + self.visual.to(torch.get_default_dtype()) self.rope_deltas = None # cache rope_deltas here self.config = config # Initialize weights and apply final processing @@ -997,41 +997,6 @@ class Qwen2_5_VLModel(nn.Module): return output if return_dict else output.to_tuple() -class DotDict(dict): - def __init__(self, mapping): - super().__init__() - for key, value in mapping.items(): - if isinstance(value, dict): - value = DotDict(value) # 递归转换 - elif isinstance(value, list): - # 如果是 list，且元素是 dict 也递归转换 - value = [ - DotDict(item) if isinstance(item, dict) else item for item in value - ] - self[key] = value - - def __getattr__(self, item): - try: - return self[item] - except KeyError: - raise AttributeError(f"No attribute '{item}'") - - def __setattr__(self, key, value): - self[key] = value - - def __delattr__(self, key): - del self[key] - - -def dict_to_namespace(d): - for k, v in d.items(): - if isinstance(v, dict): - d[k] = dict_to_namespace(v) - elif isinstance(v, list): - d[k] = [dict_to_namespace(i) if isinstance(i, dict) else i for i in v] - return SimpleNamespace(**d) - - class Qwen2_5_VLForConditionalGeneration(TextEncoder): # BitandBytes specific attributes default_bitsandbytes_target_modules = [ @@ -1058,12 +1023,17 @@ class Qwen2_5_VLForConditionalGeneration(TextEncoder): prefix: str = "", ) -> None: super().__init__(config) + enable_image_understanding = config.enable_image_understanding config = config.arch_config - self.model = Qwen2_5_VLModel(config) + self.model = Qwen2_5_VLModel( + config, enable_image_understanding=enable_image_understanding + ) self.lm_head = nn.Linear( config.text_config.hidden_size, config.text_config.vocab_size, bias=False ) + self.enable_image_understanding = enable_image_understanding + self.config = config def get_input_embeddings(self): @@ -1157,6 +1127,8 @@ class Qwen2_5_VLForConditionalGeneration(TextEncoder): name = name.replace("model.", "model.language_model.") if "visual." in name: + if not self.enable_image_understanding: + continue name = name.replace("visual.", "model.visual.") try: # Skip loading extra bias for GPTQ models. @@ -1164,7 +1136,6 @@ class Qwen2_5_VLForConditionalGeneration(TextEncoder): continue param = params_dict[name] except KeyError: - print(params_dict.keys()) raise weight_loader = getattr(param, "weight_loader", default_weight_loader) diff --git a/python/sglang/multimodal_gen/test/run_suite.py b/python/sglang/multimodal_gen/test/run_suite.py index da9eb2667..dd5ece7b5 100644 --- a/python/sglang/multimodal_gen/test/run_suite.py +++ b/python/sglang/multimodal_gen/test/run_suite.py @@ -90,7 +90,7 @@ def collect_test_items(files, filter_expr=None): cmd.extend(["-k", filter_expr]) cmd.extend(files) - logger.info(f"Collecting tests with command: {' '.join(cmd)}") + print(f"Collecting tests with command: {' '.join(cmd)}") result = subprocess.run(cmd, capture_output=True, text=True) # Check for collection errors @@ -114,7 +114,7 @@ def collect_test_items(files, filter_expr=None): raise RuntimeError(error_msg) if result.returncode == 5: - logger.info( + print( "No tests were collected (exit code 5). This may be expected with filters." ) @@ -130,7 +130,7 @@ def collect_test_items(files, filter_expr=None): if "::" in test_id: test_items.append(test_id) - logger.info(f"Collected {len(test_items)} test items") + print(f"Collected {len(test_items)} test items") return test_items @@ -145,20 +145,21 @@ def run_pytest(files, filter_expr=None): if filter_expr: base_cmd.extend(["-k", filter_expr]) - max_retries = 4 + max_retries = 6 # retry if the perf assertion failed, for {max_retries} times for i in range(max_retries + 1): cmd = list(base_cmd) if i > 0: cmd.append("--last-failed") - cmd.extend(files) + else: + cmd.extend(files) if i > 0: - logger.info( + print( f"Performance assertion failed. Retrying ({i}/{max_retries}) with --last-failed..." ) - logger.info(f"Running command: {' '.join(cmd)}") + print(f"Running command: {' '.join(cmd)}") process = subprocess.Popen( cmd, @@ -185,7 +186,7 @@ def run_pytest(files, filter_expr=None): # Exit code 5 means no tests were collected/selected - treat as success # when using filters, since some partitions may have all tests filtered out if returncode == 5: - logger.info( + print( "No tests collected (exit code 5). This is expected when filters " "deselect all tests in a partition. Treating as success." ) @@ -210,7 +211,7 @@ def run_pytest(files, filter_expr=None): if not (is_perf_assertion or is_flaky_ci_assertion or is_oom_error): return returncode - logger.info(f"Max retry exceeded") + print(f"Max retry exceeded") return returncode @@ -261,12 +262,13 @@ def main(): print(f"Selected {len(suite_files_abs)} files:") for f in suite_files_abs: print(f" - {os.path.basename(f)}") - print(f"Running {len(my_items)} items in this shard: {', '.join(my_items)}") if not my_items: print("No items assigned to this partition. Exiting success.") sys.exit(0) + print(f"Running {len(my_items)} items in this shard: {', '.join(my_items)}") + # 4. execute with the specific test items exit_code = run_pytest(my_items) sys.exit(exit_code)