opensportslib 0.2.0.dev6__tar.gz → 0.2.0.dev7__tar.gz

{opensportslib-0.2.0.dev6/opensportslib.egg-info → opensportslib-0.2.0.dev7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: opensportslib
-Version: 0.2.0.dev6
+Version: 0.2.0.dev7
 Summary: OpenSportsLib is the professional library, designed for advanced video understanding in sports. It provides state-of-the-art tools for action recognition, spotting, retrieval, and captioning, making it ideal for researchers, analysts, and developers working with sports video data.
 Author: Jeet Vora
 Requires-Python: >=3.12
@@ -26,7 +26,7 @@ Requires-Dist: tabulate
 Requires-Dist: pytorch-lightning
 Requires-Dist: pandas
 Requires-Dist: pyarrow
-Requires-Dist: h5py
+Requires-Dist: h5py==3.16.0
 Requires-Dist: huggingface_hub
 Requires-Dist: easydict
 Provides-Extra: test
@@ -317,10 +317,25 @@ single_prediction = my_model.infer(
 ```
 
 Use `opensportslib/configs/vqa/xvars.yaml` with `opensportslib setup --vqa_xvars`
-for the X-VARS-compatible backend, or `opensportslib/configs/vqa/qwen.yaml` with
-`opensportslib setup --vqa_qwen` for the Qwen-compatible backend. The Qwen
-backend currently supports `Qwen/Qwen2.5-7B-Instruct` and
-`Qwen/Qwen3.5-9B-Base`.
+for the X-VARS backend. OpenSportsLib supports three VQA options:
+
+- `opensportslib/configs/vqa/xvars.yaml`
+  Original X-VARS / Video-ChatGPT path.
+- CLIP features + Qwen
+  Use `opensportslib/configs/vqa/qwen.yaml` for inference and
+  `opensportslib/configs/vqa/qwen_lora.yaml` for LoRA training.
+- `opensportslib/configs/vqa/qwen3_vl_native.yaml`
+  Full end-to-end native QwenVL path. This is the single canonical QwenVL
+  config; change `MODEL.components.llm_decoder.params.repo_id` to switch model
+  IDs.
+
+Use `opensportslib setup --vqa_qwen` for both the CLIP+Qwen and native QwenVL
+paths. The CLIP+Qwen configs support `Qwen/Qwen2.5-7B-Instruct` and
+`Qwen/Qwen3.5-9B-Base`. The native QwenVL config defaults to
+`Qwen/Qwen3-VL-8B-Instruct` and supports:
+
+- `Qwen/Qwen3-VL-8B-Instruct`
+- `Qwen/Qwen2.5-VL-7B-Instruct`
 
 For X-VARS, `feature_source: indexed_or_raw_clip` prefers indexed CLIP features
 when available and falls back to extracting CLIP features from raw video during

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7}/README.md

@@ -281,10 +281,25 @@ single_prediction = my_model.infer(
 ```
 
 Use `opensportslib/configs/vqa/xvars.yaml` with `opensportslib setup --vqa_xvars`
-for the X-VARS-compatible backend, or `opensportslib/configs/vqa/qwen.yaml` with
-`opensportslib setup --vqa_qwen` for the Qwen-compatible backend. The Qwen
-backend currently supports `Qwen/Qwen2.5-7B-Instruct` and
-`Qwen/Qwen3.5-9B-Base`.
+for the X-VARS backend. OpenSportsLib supports three VQA options:
+
+- `opensportslib/configs/vqa/xvars.yaml`
+  Original X-VARS / Video-ChatGPT path.
+- CLIP features + Qwen
+  Use `opensportslib/configs/vqa/qwen.yaml` for inference and
+  `opensportslib/configs/vqa/qwen_lora.yaml` for LoRA training.
+- `opensportslib/configs/vqa/qwen3_vl_native.yaml`
+  Full end-to-end native QwenVL path. This is the single canonical QwenVL
+  config; change `MODEL.components.llm_decoder.params.repo_id` to switch model
+  IDs.
+
+Use `opensportslib setup --vqa_qwen` for both the CLIP+Qwen and native QwenVL
+paths. The CLIP+Qwen configs support `Qwen/Qwen2.5-7B-Instruct` and
+`Qwen/Qwen3.5-9B-Base`. The native QwenVL config defaults to
+`Qwen/Qwen3-VL-8B-Instruct` and supports:
+
+- `Qwen/Qwen3-VL-8B-Instruct`
+- `Qwen/Qwen2.5-VL-7B-Instruct`
 
 For X-VARS, `feature_source: indexed_or_raw_clip` prefers indexed CLIP features
 when available and falls back to extracting CLIP features from raw video during

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7}/opensportslib/apis/localization.py

@@ -1,6 +1,7 @@
 import logging
 import os
 import time
+from types import SimpleNamespace
 
 from opensportslib.apis.base_task_model import BaseTaskModel
 from opensportslib.core.config.accessors import (
@@ -14,6 +15,7 @@ from opensportslib.core.config.accessors import (
     get_split_annotation_path,
     get_split_cfg,
     set_split_annotation_path,
+    set_loader_backend,
     get_model_family,
 )
 from opensportslib.core.utils.config import expand
@@ -22,6 +24,17 @@ from opensportslib.core.utils.config import expand
 class LocalizationModel(BaseTaskModel):
     """Top-level task wrapper for localization / spotting."""
 
+    _HF_BACKEND_SPLIT_TYPES = {
+        "dali": {
+            "VideoGameWithOpencv": "VideoGameWithDali",
+            "VideoGameWithOpencvVideo": "VideoGameWithDaliVideo",
+        },
+        "opencv": {
+            "VideoGameWithDali": "VideoGameWithOpencv",
+            "VideoGameWithDaliVideo": "VideoGameWithOpencvVideo",
+        },
+    }
+
     # def __init__(self, config=None, weights=None):
     #     super().__init__(config=config, weights=None)
     #     if weights is not None:
@@ -62,6 +75,82 @@ class LocalizationModel(BaseTaskModel):
                 device,
             )
 
+    @staticmethod
+    def _device_type(device) -> str:
+        device_type = getattr(device, "type", device)
+        return str(device_type).split(":", 1)[0].lower()
+
+    @staticmethod
+    def _is_hf_repo_weights(weights: str | None) -> bool:
+        if not weights:
+            return False
+        from opensportslib.core.utils.config import is_local_path
+
+        return not is_local_path(weights)
+
+    @staticmethod
+    def _iter_split_items(splits):
+        if splits is None:
+            return []
+        if isinstance(splits, dict):
+            return list(splits.items())
+        return list(vars(splits).items())
+
+    @staticmethod
+    def _ensure_namespace(parent, attr: str):
+        current = getattr(parent, attr, None)
+        if current is None:
+            current = SimpleNamespace()
+            setattr(parent, attr, current)
+        return current
+
+    def _normalize_opencv_dataloader(self, split_name: str, split_cfg) -> None:
+        dataloader = getattr(split_cfg, "dataloader", None)
+        if dataloader is None:
+            dataloader = SimpleNamespace()
+            setattr(split_cfg, "dataloader", dataloader)
+
+        if getattr(dataloader, "batch_size", None) is None:
+            dataloader.batch_size = 1
+        if getattr(dataloader, "shuffle", None) is None:
+            dataloader.shuffle = split_name == "train"
+        if getattr(dataloader, "num_workers", None) is None:
+            dataloader.num_workers = 0
+        if getattr(dataloader, "pin_memory", None) is None:
+            dataloader.pin_memory = False
+
+    def _adapt_hf_backend_for_device(self, weights: str | None) -> None:
+        if not self._is_hf_repo_weights(weights):
+            return
+
+        from opensportslib.core.utils.config import select_device
+
+        device = select_device(self.config.SYSTEM)
+        backend = "dali" if self._device_type(device) == "cuda" else "opencv"
+        current_backend = get_loader_backend(self.config)
+        set_loader_backend(self.config, backend)
+
+        common = getattr(getattr(self.config, "DATA", None), "common", None)
+        if common is None:
+            return
+        splits = self._ensure_namespace(common, "splits")
+        remap = self._HF_BACKEND_SPLIT_TYPES[backend]
+
+        for split_name, split_cfg in self._iter_split_items(splits):
+            split_type = getattr(split_cfg, "type", None)
+            if split_type in remap:
+                setattr(split_cfg, "type", remap[split_type])
+            if backend == "opencv":
+                self._normalize_opencv_dataloader(split_name, split_cfg)
+
+        if current_backend != backend:
+            logging.info(
+                "HF localization runtime backend override: %s -> %s for weights=%s",
+                current_backend,
+                backend,
+                weights,
+            )
+
     def load_weights(
         self,
         weights: str | None = None,
@@ -80,6 +169,7 @@ class LocalizationModel(BaseTaskModel):
         if weights is None:
             raise ValueError("`weights` must be provided to load_weights().")
 
+        self._adapt_hf_backend_for_device(weights)
         device = select_device(self.config.SYSTEM)
         self._gate_multi_gpu_by_device(device)
         if self.model is None:
@@ -183,8 +273,10 @@ class LocalizationModel(BaseTaskModel):
         # E2E validation mAP uses the `valid_data_frames` split; keep it in sync
         # with explicit valid annotation overrides.
         self._set_split_path("valid_data_frames", valid_set)
-        
+
         self.config = resolve_config_omega(self.config, weights=weights)
+        effective_weights = weights if weights is not None else self.last_loaded_weights
+        self._adapt_hf_backend_for_device(effective_weights)
         check_config(self.config, split="train")
         init_wandb(
             self.config_path,
@@ -196,8 +288,6 @@ class LocalizationModel(BaseTaskModel):
         logging.info("Configuration:")
         logging.info(self.config)
 
-        effective_weights = weights if weights is not None else self.last_loaded_weights
-
         def set_seed(seed):
             random.seed(seed)
             np.random.seed(seed)
@@ -298,6 +388,8 @@ class LocalizationModel(BaseTaskModel):
         self._set_split_path("test", test_set)
 
         self.config = resolve_config_omega(self.config, weights=weights)
+        effective_weights = weights if weights is not None else self.last_loaded_weights
+        self._adapt_hf_backend_for_device(effective_weights)
         check_config(self.config, split="test")
         self.config.infer_split = whether_infer_split(get_split_cfg(self.config, "test"))
 
@@ -313,8 +405,6 @@ class LocalizationModel(BaseTaskModel):
 
         start = time.time()
 
-        effective_weights = weights if weights is not None else self.last_loaded_weights
-
         if effective_weights is not None:
             if self.model is None or self.last_loaded_weights != effective_weights:
                 self.load_weights(weights=effective_weights)
@@ -368,6 +458,8 @@ class LocalizationModel(BaseTaskModel):
         test_set = self._resolve_split_path("test", test_set)
         self._set_split_path("test", test_set)
         self.config = resolve_config_omega(self.config, weights=weights)
+        effective_weights = weights if weights is not None else self.last_loaded_weights
+        self._adapt_hf_backend_for_device(effective_weights)
         check_config(self.config, split="test")
         self.config.infer_split = whether_infer_split(get_split_cfg(self.config, "test"))
 

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7}/opensportslib/configs/default.yaml

@@ -2,7 +2,7 @@ SYSTEM:
   paths:
     log_dir: ./logs
     save_dir: ./checkpoints
-  device: cuda
+  device: auto
   gpu:
     id: 0
   reproducibility:

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7}/opensportslib/configs/vqa/default.yaml

@@ -5,7 +5,7 @@ SYSTEM:
   paths:
     save_dir: ./checkpoints_vqa
     work_dir: ${SYSTEM.paths.save_dir}
-  device: cuda
+  device: auto
   gpu:
     count: 1
     id: 0

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7}/opensportslib/configs/vqa/qwen3_vl_native.yaml

@@ -25,8 +25,15 @@ MODEL:
       kind: decoder
       source:
         provider: huggingface
-        name:  Qwen/Qwen3-VL-8B-Instruct
+        # Canonical end-to-end QwenVL config.
+        # Swap this model ID if you want another supported QwenVL checkpoint.
+        # - Qwen/Qwen3-VL-8B-Instruct
+        # - Qwen/Qwen2.5-VL-7B-Instruct
+        name: Qwen/Qwen3-VL-8B-Instruct
       params:
+        # Supported examples:
+        # - Qwen/Qwen3-VL-8B-Instruct
+        # - Qwen/Qwen2.5-VL-7B-Instruct
         repo_id: Qwen/Qwen3-VL-8B-Instruct
       overrides: {}
   metadata:

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7}/opensportslib/core/config/accessors.py

@@ -68,6 +68,20 @@ def get_loader_backend(cfg: Any) -> str:
     return str(backend).lower()
 
 
+def set_loader_backend(cfg: Any, backend: str) -> None:
+    data = getattr(cfg, "DATA", None)
+    if data is None:
+        data = SimpleNamespace()
+        setattr(cfg, "DATA", data)
+
+    common = _ensure_child(data, "common")
+    runtime = _ensure_child(common, "runtime")
+    if isinstance(runtime, dict):
+        runtime["loader_backend"] = str(backend).lower()
+    else:
+        setattr(runtime, "loader_backend", str(backend).lower())
+
+
 def get_system_path(cfg: Any, key: str, default: str | None = None) -> str | None:
     system = _as_dict(getattr(cfg, "SYSTEM", None))
     paths = _as_dict(system.get("paths"))
@@ -351,6 +365,30 @@ def get_train_execution(cfg: Any) -> dict[str, Any]:
     return _as_dict(train.get("execution"))
 
 
+def get_hf_prefer_cuda(cfg: Any, hf_cfg: dict[str, Any] | None = None) -> bool:
+    """Resolve whether HF-backed runtimes should prefer CUDA.
+
+    Precedence:
+    1. explicit TRAIN.execution.hf.prefer_cuda
+    2. SYSTEM.device mapping:
+       - cpu -> False
+       - cuda/gpu -> True
+       - auto/unset -> True
+    """
+    hf_cfg = _as_dict(hf_cfg) if hf_cfg is not None else _as_dict(get_train_execution(cfg).get("hf"))
+    explicit = hf_cfg.get("prefer_cuda")
+    if explicit is not None:
+        return bool(explicit)
+
+    system = _as_dict(getattr(cfg, "SYSTEM", None))
+    mode = str(system.get("device", "auto")).strip().lower()
+    if mode == "cpu":
+        return False
+    if mode in {"cuda", "gpu"}:
+        return True
+    return True
+
+
 def get_hf_cuda_device_index(cfg: Any, hf_cfg: dict[str, Any] | None = None) -> int | None:
     """Resolve the CUDA device index used by HuggingFace runtime helpers."""
     if os.environ.get("CUDA_VISIBLE_DEVICES"):

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7}/opensportslib/core/trainer/vqa_trainer.py

@@ -17,24 +17,25 @@ import torch
 from tqdm.auto import tqdm
 
 from opensportslib.core.config.accessors import (
-    get_xvars_train_model_id,
-    get_xvars_train_tokenizer_id,
-    get_xvars_train_video_token_len,
     get_hf_cuda_device_index,
-    get_vqa_backend,
-    get_vqa_decoder_model_id,
-    is_xvars_videochatgpt_backend,
+    get_hf_prefer_cuda,
     get_model_runtime_dtype,
     get_split_dataloader_cfg,
     get_system_path,
     get_train_epochs,
     get_train_execution,
     get_train_optimizer,
+    get_vqa_backend,
+    get_vqa_decoder_model_id,
     get_vqa_eval_profile_cfg,
     get_vqa_generation_cfg,
     get_vqa_mm_hidden_size,
     get_vqa_native_visual_cfg,
     get_vqa_prompt_video_token_len,
+    get_xvars_train_model_id,
+    get_xvars_train_tokenizer_id,
+    get_xvars_train_video_token_len,
+    is_xvars_videochatgpt_backend,
 )
 from opensportslib.core.utils.config import save_config
 from opensportslib.core.utils.hf_runtime import (
@@ -952,7 +953,7 @@ class VQAXVarsVideoChatGPTLoraTrainer:
         cuda_device_index = get_hf_cuda_device_index(self.config, hf_cfg)
         if bnb_config is not None:
             model_kwargs["quantization_config"] = bnb_config
-            if bool(hf_cfg.get("prefer_cuda", True)) and torch.cuda.is_available():
+            if get_hf_prefer_cuda(self.config, hf_cfg) and torch.cuda.is_available():
                 if cuda_device_index is not None:
                     torch.cuda.set_device(cuda_device_index)
                     model_kwargs["device_map"] = {"": cuda_device_index}
@@ -1033,7 +1034,7 @@ class VQAXVarsVideoChatGPTLoraTrainer:
             "report_to": ["wandb"] if use_wandb else [],
             "remove_unused_columns": False,
             "disable_tqdm": bool(sft_cfg.get("disable_tqdm", True)),
-            "use_cpu": not bool(hf_cfg.get("prefer_cuda", True)),
+            "use_cpu": not get_hf_prefer_cuda(self.config, hf_cfg),
             "fp16": fp16,
             "bf16": bf16,
             "gradient_checkpointing": bool(sft_cfg.get("gradient_checkpointing", False)),
@@ -1186,7 +1187,7 @@ class VQAQwenXVarsLoraTrainer:
         cuda_device_index = get_hf_cuda_device_index(self.config, hf_cfg)
         if bnb_config is not None:
             model_kwargs["quantization_config"] = bnb_config
-            if bool(hf_cfg.get("prefer_cuda", True)) and torch.cuda.is_available():
+            if get_hf_prefer_cuda(self.config, hf_cfg) and torch.cuda.is_available():
                 if cuda_device_index is not None:
                     torch.cuda.set_device(cuda_device_index)
                     model_kwargs["device_map"] = {"": cuda_device_index}
@@ -1261,7 +1262,7 @@ class VQAQwenXVarsLoraTrainer:
             "report_to": ["wandb"] if use_wandb else [],
             "remove_unused_columns": False,
             "disable_tqdm": bool(sft_cfg.get("disable_tqdm", True)),
-            "use_cpu": not bool(hf_cfg.get("prefer_cuda", True)),
+            "use_cpu": not get_hf_prefer_cuda(self.config, hf_cfg),
             "fp16": fp16,
             "bf16": bf16,
             "gradient_checkpointing": bool(sft_cfg.get("gradient_checkpointing", False)),
@@ -1491,7 +1492,7 @@ class VQAQwenVLNativeLoraTrainer:
             "report_to": ["wandb"] if use_wandb else [],
             "remove_unused_columns": False,
             "disable_tqdm": bool(sft_cfg.get("disable_tqdm", True)),
-            "use_cpu": not bool(hf_cfg.get("prefer_cuda", True)),
+            "use_cpu": not get_hf_prefer_cuda(self.config, hf_cfg),
             "fp16": fp16,
             "bf16": bf16,
             "gradient_checkpointing": bool(sft_cfg.get("gradient_checkpointing", False)),

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7}/opensportslib/models/base/qwen_vl_native.py

@@ -13,6 +13,7 @@ import torch.nn as nn
 
 from opensportslib.core.config.accessors import (
     get_hf_cuda_device_index,
+    get_hf_prefer_cuda,
     get_model_load,
     get_train_execution,
     get_vqa_generation_cfg,
@@ -248,7 +249,7 @@ class QwenVLNativeModel(nn.Module):
         hf_cfg = _as_dict(exec_cfg.get("hf"))
         quant_cfg = _as_dict(exec_cfg.get("quantization"))
         local_files_only = bool(hf_cfg.get("local_files_only", False))
-        prefer_cuda = bool(hf_cfg.get("prefer_cuda", True))
+        prefer_cuda = get_hf_prefer_cuda(config, hf_cfg)
         cuda_device_index = get_hf_cuda_device_index(config, hf_cfg)
         use_cuda = prefer_cuda and torch.cuda.is_available()
         if use_cuda and cuda_device_index is not None:

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7}/opensportslib/models/base/qwen_xvars.py

@@ -13,6 +13,7 @@ from opensportslib.core.config.accessors import (
     get_component_load_by_kind,
     get_component_params_by_kind,
     get_hf_cuda_device_index,
+    get_hf_prefer_cuda,
     get_model_load,
     get_model_runtime_dtype,
     get_train_execution,
@@ -363,7 +364,7 @@ class QwenXVarsModel(nn.Module):
         )
 
         local_files_only = bool(hf_cfg.get("local_files_only", False))
-        prefer_cuda = bool(hf_cfg.get("prefer_cuda", True))
+        prefer_cuda = get_hf_prefer_cuda(config, hf_cfg)
         cuda_device_index = get_hf_cuda_device_index(config, hf_cfg)
         mm_hidden_size = int(projector_params.get("input_dim") or get_vqa_mm_hidden_size(config, default=1024))
         use_cuda = prefer_cuda and torch.cuda.is_available()
@@ -481,7 +482,7 @@ class QwenXVarsModel(nn.Module):
                         self.raw_extractor = XVarsStrictRawVideoFeatureExtractor(
                             weights_path=self.vision_weights_path,
                             vision_tower=self.vision_tower_name,
-                            prefer_cuda=bool(hf_cfg.get("prefer_cuda", True)),
+                            prefer_cuda=get_hf_prefer_cuda(config, hf_cfg),
                             start_frame=self.strict_sampling_cfg.get("start_frame"),
                             end_frame=self.strict_sampling_cfg.get("end_frame"),
                             input_fps=self.strict_sampling_cfg.get("input_fps"),
@@ -491,7 +492,7 @@ class QwenXVarsModel(nn.Module):
                     else:
                         self.raw_extractor = XVarsRawVideoFeatureExtractor(
                             vision_tower=self.vision_tower_name,
-                            prefer_cuda=bool(hf_cfg.get("prefer_cuda", True)),
+                            prefer_cuda=get_hf_prefer_cuda(config, hf_cfg),
                         )
                 if isinstance(self.raw_extractor, XVarsStrictRawVideoFeatureExtractor):
                     features, classifier_prior = self.raw_extractor.extract_with_prior(video_path)

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7}/opensportslib/models/base/xvars_videochatgpt.py

@@ -19,16 +19,17 @@ import torch.nn as nn
 from opensportslib.core.config.accessors import (
     get_component_load_by_kind,
     get_component_params_by_kind,
-    get_hf_cuda_device_index,
     get_data_sampling,
-    get_xvars_infer_tokenizer_id,
-    get_xvars_infer_video_token_len,
+    get_hf_cuda_device_index,
+    get_hf_prefer_cuda,
     get_model_load,
     get_model_runtime_dtype,
     get_train_execution,
     get_vqa_feature_source,
     get_vqa_mm_hidden_size,
     get_vqa_xvars_feature_mode,
+    get_xvars_infer_tokenizer_id,
+    get_xvars_infer_video_token_len,
 )
 from opensportslib.core.utils.hf_runtime import (
     VIDEO_SPECIAL_TOKENS,
@@ -856,7 +857,7 @@ class XVarsVideoChatGPTModel(nn.Module):
         )
 
         local_files_only = bool(hf_cfg.get("local_files_only", False))
-        prefer_cuda = bool(hf_cfg.get("prefer_cuda", True))
+        prefer_cuda = get_hf_prefer_cuda(config, hf_cfg)
         cuda_device_index = get_hf_cuda_device_index(config, hf_cfg)
         adapter_path = get_model_load(config).get("checkpoint_path")
         projection_path = xvars_cfg.get("projection_path")
@@ -998,7 +999,7 @@ class XVarsVideoChatGPTModel(nn.Module):
                         self.raw_extractor = XVarsStrictRawVideoFeatureExtractor(
                             weights_path=self.vision_weights_path,
                             vision_tower=self.vision_tower_name,
-                            prefer_cuda=bool(hf_cfg.get("prefer_cuda", True)),
+                            prefer_cuda=get_hf_prefer_cuda(self.config, hf_cfg),
                             start_frame=self.strict_sampling_cfg.get("start_frame"),
                             end_frame=self.strict_sampling_cfg.get("end_frame"),
                             input_fps=self.strict_sampling_cfg.get("input_fps"),
@@ -1008,7 +1009,7 @@ class XVarsVideoChatGPTModel(nn.Module):
                     else:
                         self.raw_extractor = XVarsRawVideoFeatureExtractor(
                             vision_tower=self.vision_tower_name,
-                            prefer_cuda=bool(hf_cfg.get("prefer_cuda", True)),
+                            prefer_cuda=get_hf_prefer_cuda(self.config, hf_cfg),
                         )
                 if isinstance(self.raw_extractor, XVarsStrictRawVideoFeatureExtractor):
                     logger.info("USING STRICT XVARS EXTRACTOR")

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7/opensportslib.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: opensportslib
-Version: 0.2.0.dev6
+Version: 0.2.0.dev7
 Summary: OpenSportsLib is the professional library, designed for advanced video understanding in sports. It provides state-of-the-art tools for action recognition, spotting, retrieval, and captioning, making it ideal for researchers, analysts, and developers working with sports video data.
 Author: Jeet Vora
 Requires-Python: >=3.12
@@ -26,7 +26,7 @@ Requires-Dist: tabulate
 Requires-Dist: pytorch-lightning
 Requires-Dist: pandas
 Requires-Dist: pyarrow
-Requires-Dist: h5py
+Requires-Dist: h5py==3.16.0
 Requires-Dist: huggingface_hub
 Requires-Dist: easydict
 Provides-Extra: test
@@ -317,10 +317,25 @@ single_prediction = my_model.infer(
 ```
 
 Use `opensportslib/configs/vqa/xvars.yaml` with `opensportslib setup --vqa_xvars`
-for the X-VARS-compatible backend, or `opensportslib/configs/vqa/qwen.yaml` with
-`opensportslib setup --vqa_qwen` for the Qwen-compatible backend. The Qwen
-backend currently supports `Qwen/Qwen2.5-7B-Instruct` and
-`Qwen/Qwen3.5-9B-Base`.
+for the X-VARS backend. OpenSportsLib supports three VQA options:
+
+- `opensportslib/configs/vqa/xvars.yaml`
+  Original X-VARS / Video-ChatGPT path.
+- CLIP features + Qwen
+  Use `opensportslib/configs/vqa/qwen.yaml` for inference and
+  `opensportslib/configs/vqa/qwen_lora.yaml` for LoRA training.
+- `opensportslib/configs/vqa/qwen3_vl_native.yaml`
+  Full end-to-end native QwenVL path. This is the single canonical QwenVL
+  config; change `MODEL.components.llm_decoder.params.repo_id` to switch model
+  IDs.
+
+Use `opensportslib setup --vqa_qwen` for both the CLIP+Qwen and native QwenVL
+paths. The CLIP+Qwen configs support `Qwen/Qwen2.5-7B-Instruct` and
+`Qwen/Qwen3.5-9B-Base`. The native QwenVL config defaults to
+`Qwen/Qwen3-VL-8B-Instruct` and supports:
+
+- `Qwen/Qwen3-VL-8B-Instruct`
+- `Qwen/Qwen2.5-VL-7B-Instruct`
 
 For X-VARS, `feature_source: indexed_or_raw_clip` prefers indexed CLIP features
 when available and falls back to extracting CLIP features from raw video during

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7}/opensportslib.egg-info/SOURCES.txt

@@ -39,8 +39,6 @@ opensportslib/configs/vqa/qwen.yaml
 opensportslib/configs/vqa/qwen3_vl_native.yaml
 opensportslib/configs/vqa/qwen_lora.yaml
 opensportslib/configs/vqa/qwen_sngar_frames.yaml
-opensportslib/configs/vqa/qwen_vl_native.yaml
-opensportslib/configs/vqa/qwen_vl_native_awq.yaml
 opensportslib/configs/vqa/xvars.yaml
 opensportslib/core/__init__.py
 opensportslib/core/config/__init__.py
@@ -175,4 +173,5 @@ tools/download/download_osl_hf.py
 tools/download/upload_osl_hf.py
 tools/training/classification.py
 tools/training/localization.py
-tools/training/vqa.py
+tools/training/vqa.py
+tools/upload/upload_model_hf.py

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7}/opensportslib.egg-info/requires.txt

@@ -16,7 +16,7 @@ tabulate
 pytorch-lightning
 pandas
 pyarrow
-h5py
+h5py==3.16.0
 huggingface_hub
 easydict
 

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7}/pyproject.toml

@@ -4,11 +4,11 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "opensportslib"
-version = "0.2.0.dev6"
+version = "0.2.0.dev7"
 description = "OpenSportsLib is the professional library, designed for advanced video understanding in sports. It provides state-of-the-art tools for action recognition, spotting, retrieval, and captioning, making it ideal for researchers, analysts, and developers working with sports video data."
 readme = "README.md"
 requires-python = ">=3.12"
-dependencies = [ "SoccerNet", "av", "decord; platform_system != 'Darwin' and platform_machine != 'arm64' and platform_machine != 'aarch64'", "evaluate", "scikit-learn", "torch", "torchvision", "transformers==4.57.3", "tokenizers==0.22.1", "accelerate", "wandb", "opencv-python", "omegaconf", "timm", "seaborn", "tabulate", "pytorch-lightning", "pandas", "pyarrow", "h5py", "huggingface_hub", "easydict",]
+dependencies = [ "SoccerNet", "av", "decord; platform_system != 'Darwin' and platform_machine != 'arm64' and platform_machine != 'aarch64'", "evaluate", "scikit-learn", "torch", "torchvision", "transformers==4.57.3", "tokenizers==0.22.1", "accelerate", "wandb", "opencv-python", "omegaconf", "timm", "seaborn", "tabulate", "pytorch-lightning", "pandas", "pyarrow", "h5py==3.16.0", "huggingface_hub", "easydict",]
 [[project.authors]]
 name = "Jeet Vora"
 

{opensportslib-0.2.0.dev6 → opensportslib-0.2.0.dev7}/tests/test_config_architecture.py

@@ -112,6 +112,20 @@ def test_vqa_qwen_experiment_composes_all_layers():
     assert cfg["TRAIN"]["execution"]["hf"]["offload_folder"] == "./hf_offload_qwen"
 
 
+def test_vqa_qwen_vl_native_experiment_composes_all_layers():
+    cfg = load_config("opensportslib/configs/vqa/qwen3_vl_native.yaml", as_namespace=False)
+
+    assert cfg["VERSION"] == 2
+    assert cfg["TASK"] == "vqa"
+    assert cfg["SYSTEM"]["paths"]["log_dir"] == "./logs"
+    assert cfg["SYSTEM"]["paths"]["save_dir"] == "./checkpoints_vqa_qwen3_vl_native"
+    assert cfg["SYSTEM"]["paths"]["work_dir"] == "./checkpoints_vqa_qwen3_vl_native"
+    assert cfg["DATA"]["common"]["runtime"]["loader_backend"] == "opencv"
+    assert cfg["MODEL"]["metadata"]["backend"] == "qwen_vl_native_infer"
+    assert cfg["MODEL"]["components"]["llm_decoder"]["source"]["name"] == "Qwen/Qwen3-VL-8B-Instruct"
+    assert cfg["TRAIN"]["execution"]["training_backend"] == "qwen_vl_native_lora"
+
+
 def test_validation_accepts_canonical_schema():
     canonical = load_config(
         "opensportslib/configs/localization/default.yaml",