Port to transformers 5.x API (hub 1.x / gradio 6 compatible)

README.md

@@ -47,7 +47,9 @@ repo, and message construction, the duration grid, reference labelling /
 validation and the output schema check are ported 1:1 from the adapter repo's
 `infer.py`. Encoding uses the same `ProcessorMixin.apply_chat_template` path
 with `max_pixels` 301056 (images) / 100352 (video), `load_audio_from_video=False`
-and `use_audio_in_video=False`.
+and `use_audio_in_video=False`. Those processing kwargs are passed through
+transformers 5.x's `processor_kwargs` dict (transformers 4.x, which the
+reference pins, is incompatible with `huggingface_hub` 1.x / Gradio 6).
 
 Runs on ZeroGPU: the Thinker is loaded in bf16 with SDPA attention and the LoRA
 is attached with PEFT at module scope.

app.py

@@ -33,6 +33,7 @@ import gradio as gr
 from huggingface_hub import hf_hub_download
 from safetensors import safe_open
 from transformers import (
+    Qwen2_5OmniConfig,
     Qwen2_5OmniProcessor,
     Qwen2_5OmniThinkerForConditionalGeneration,
     TextIteratorStreamer,
@@ -46,6 +47,9 @@ ADAPTER_REPO = "lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA-Omni"
 # Defaults taken from the adapter repo's infer.py CLI.
 IMAGE_MAX_PIXELS = 301056
 VIDEO_MAX_PIXELS = 100352
+# Qwen2-VL image-processor default lower bound, kept explicit because
+# transformers 5.x only honours max_pixels when min_pixels is given too.
+IMAGE_MIN_PIXELS = 56 * 56
 
 TASKS = ["T2AV", "I2AV", "L2AV", "FL2AV", "Ref2AV"]
 RATIOS = ["adaptive", "21:9", "16:9", "4:3", "1:1", "3:4", "9:16"]
@@ -265,22 +269,25 @@ def output_schema_ok(task: str, text: str) -> bool:
 
 print(f"[boot] Loading Qwen2.5-Omni processor from {BASE_MODEL} …", flush=True)
 processor = Qwen2_5OmniProcessor.from_pretrained(BASE_MODEL, trust_remote_code=False)
-processor.image_processor.max_pixels = IMAGE_MAX_PIXELS
-try:
-    processor.video_processor.max_pixels = VIDEO_MAX_PIXELS
-except Exception as error:  # pragma: no cover - attribute layout varies by version
-    print(f"[boot] video_processor.max_pixels not settable ({error!r})", flush=True)
-processor.tokenizer.padding_side = "right"
 if processor.tokenizer.pad_token_id is None:
     processor.tokenizer.pad_token = processor.tokenizer.eos_token
 
 print(f"[boot] Loading Thinker weights from {BASE_MODEL} (bf16, sdpa) …", flush=True)
+# The checkpoint is the full Omni model; take the thinker sub-config explicitly
+# and let `base_model_prefix = "thinker"` strip the `thinker.` weight prefix.
+_thinker_config = Qwen2_5OmniConfig.from_pretrained(BASE_MODEL).thinker_config
 model = Qwen2_5OmniThinkerForConditionalGeneration.from_pretrained(
     BASE_MODEL,
+    config=_thinker_config,
     dtype=torch.bfloat16,
     attn_implementation="sdpa",
     trust_remote_code=False,
-    low_cpu_mem_usage=True,
+)
+print(
+    "[boot] Thinker loaded: hidden_size="
+    f"{model.config.text_config.hidden_size}, "
+    f"vocab_size={model.config.text_config.vocab_size}",
+    flush=True,
 )
 
 print(f"[boot] Applying LoRA adapter {ADAPTER_REPO} …", flush=True)
@@ -291,7 +298,7 @@ from peft import LoraConfig, PeftModel, set_peft_model_state_dict  # noqa: E402
 
 _adapter_file = hf_hub_download(ADAPTER_REPO, "adapter_model.safetensors")
 _adapter_state_dict: dict[str, torch.Tensor] = {}
-with safe_open(_adapter_file, framework="numpy") as handle:
+with safe_open(_adapter_file, framework="numpy", device="cpu") as handle:
     for key in handle.keys():
         _adapter_state_dict[key] = torch.from_numpy(handle.get_tensor(key))
 
@@ -340,15 +347,23 @@ def _encode(messages: list[dict[str, Any]], video_fps: float) -> dict[str, Any]:
         add_generation_prompt=True,
         return_dict=True,
         return_tensors="pt",
-        text_kwargs={"padding": False},
-        images_kwargs={"max_pixels": IMAGE_MAX_PIXELS},
-        videos_kwargs={
-            "max_pixels": VIDEO_MAX_PIXELS,
-            "fps": video_fps,
-            "use_audio_in_video": False,
-        },
-        video_fps=video_fps,
         load_audio_from_video=False,
+        # transformers 5.x: processing kwargs must live in `processor_kwargs`.
+        # Passing any other flat kwarg here silently REPLACES this dict.
+        processor_kwargs={
+            "text_kwargs": {"padding": False},
+            "images_kwargs": {
+                "min_pixels": IMAGE_MIN_PIXELS,
+                "max_pixels": IMAGE_MAX_PIXELS,
+            },
+            "videos_kwargs": {
+                "min_pixels": VIDEO_MAX_PIXELS,
+                "max_pixels": VIDEO_MAX_PIXELS,
+                "fps": video_fps,
+                "do_sample_frames": True,
+                "use_audio_in_video": False,
+            },
+        },
     )
     encoded = dict(encoded)
     encoded["use_audio_in_video"] = False

requirements.txt

@@ -1,9 +1,8 @@
-# Pinned to the last 4.x: transformers 5.x changed
-# ProcessorMixin.apply_chat_template to require a `processor_kwargs` dict,
-# which is incompatible with the adapter's reference encoding path.
-transformers==4.57.6
+# transformers 4.x pins huggingface-hub<1.0, which conflicts with gradio 6,
+# so this Space runs the 5.x API (processing kwargs go in `processor_kwargs`).
+transformers==5.16.1
 accelerate>=1.10
-peft>=0.15.2
+peft>=0.18
 safetensors>=0.5.0
 torchvision
 numpy