DepthLens / src /models /vlm.py
Rishabh Jain
Prepare Space for ZeroGPU deployment
842d4e9
"""
VLM loader with hardware auto-detection.
Auto-detects CUDA VRAM and loads appropriate model:
- Moondream 2B for <16GB VRAM (Colab T4)
- Qwen2.5-VL-3B for >=16GB VRAM (RTX 5060)
Addresses critical model loading bugs encountered during development.
"""
import importlib.util
from typing import Optional
import torch
from PIL import Image
from transformers import AutoModelForCausalLM, AutoProcessor
from ..config import (
MOONDREAM_ID,
MOONDREAM_REVISION,
QWEN_ID,
QWEN_MAX_PIXELS,
QWEN_MIN_PIXELS,
)
def get_vram_gb() -> float:
"""Return total CUDA VRAM in GB for device 0, or 0.0 if CUDA unavailable.
Returns:
Total VRAM in gigabytes.
"""
if not torch.cuda.is_available():
return 0.0
return torch.cuda.get_device_properties(0).total_memory / (1024 ** 3)
def _device_map() -> dict | str:
"""Pick a device_map that works on both GPU and CPU hosts."""
return {"": "cuda"} if torch.cuda.is_available() else {"": "cpu"}
def _has_flash_attention() -> bool:
"""Check whether flash-attn is installed."""
return importlib.util.find_spec("flash_attn") is not None
def _has_qwen_vl_utils() -> bool:
"""Check whether qwen_vl_utils is installed."""
return importlib.util.find_spec("qwen_vl_utils") is not None
class VLM:
"""Vision-Language Model loader with automatic hardware-based model selection.
Loads Moondream 2B on systems with <16 GB VRAM (e.g. Colab T4) and
Qwen2.5-VL-3B on systems with >=16 GB VRAM (e.g. RTX 5060).
"""
def __init__(self, force_model: Optional[str] = None) -> None:
"""Detect available VRAM and load the appropriate VLM.
Args:
force_model: Override auto-selection. Pass ``"moondream"`` or
``"qwen"`` to skip the VRAM threshold check.
"""
self.vram_gb: float = get_vram_gb()
self.model_type: str = ""
self.model: Optional[torch.nn.Module] = None
self.processor: Optional[AutoProcessor] = None
use_model = force_model or ("qwen" if self.vram_gb >= 16.0 else "moondream")
if use_model == "qwen":
self._load_qwen()
else:
self._load_moondream()
# ── Loaders ──────────────────────────────────────────────────────────────
def _load_moondream(self) -> None:
"""Load Moondream 2B (for <16 GB VRAM).
CRITICAL: device_map must be {"": "cuda"}, NOT "auto" β€” "auto" crashes
on transformers 5.x due to the all_tied_weights_keys bug.
Revision is pinned to "2025-06-21" for API stability.
"""
print(f"Loading Moondream 2B (VRAM: {self.vram_gb:.1f} GB)...")
self.model_type = "moondream"
self.model = AutoModelForCausalLM.from_pretrained(
MOONDREAM_ID,
revision=MOONDREAM_REVISION,
device_map=_device_map(),
trust_remote_code=True,
)
if torch.cuda.is_available():
print(
f" GPU memory allocated: "
f"{torch.cuda.memory_allocated() / 1024**2:.0f} MB"
)
def _load_qwen(self) -> None:
"""Load Qwen2.5-VL-3B (for >=16 GB VRAM).
CRITICAL: pixel caps MUST be set on processor.image_processor (not the
processor wrapper). Without this, dynamic resolution can generate 16k+
tokens and cause OOM on a 16 GB GPU.
"""
print(f"Loading Qwen2.5-VL-3B (VRAM: {self.vram_gb:.1f} GB)...")
self.model_type = "qwen"
self.processor = AutoProcessor.from_pretrained(QWEN_ID, use_fast=True)
# Set pixel caps on the image_processor sub-object β€” this is where
# Qwen2_5_VLImageProcessor reads them during __call__.
self.processor.image_processor.min_pixels = QWEN_MIN_PIXELS
self.processor.image_processor.max_pixels = QWEN_MAX_PIXELS
load_kwargs: dict = {
"device_map": _device_map(),
"torch_dtype": torch.float16 if torch.cuda.is_available() else torch.float32,
}
if torch.cuda.is_available() and _has_flash_attention():
load_kwargs["attn_implementation"] = "flash_attention_2"
from transformers import AutoModelForVision2Seq # only available in transformers <5
self.model = AutoModelForVision2Seq.from_pretrained(QWEN_ID, **load_kwargs)
if torch.cuda.is_available():
print(
f" GPU memory allocated: "
f"{torch.cuda.memory_allocated() / 1024**2:.0f} MB"
)
# ── Inference ─────────────────────────────────────────────────────────────
def query_vlm(self, image: Image.Image, question: str) -> str:
"""Query the loaded VLM with an image and a text question.
Args:
image: PIL Image to analyse.
question: Text question or prompt about the image.
Returns:
The model's text response (prompt stripped).
"""
if self.model_type == "moondream":
return self._query_moondream(image, question)
if self.model_type == "qwen":
return self._query_qwen(image, question)
raise RuntimeError(f"No VLM loaded (model_type={self.model_type!r})")
def _query_moondream(self, image: Image.Image, question: str) -> str:
"""Query Moondream using the new query() API.
Args:
image: PIL Image.
question: Text question.
Returns:
Answer string from model.query()["answer"].
"""
with torch.inference_mode():
return self.model.query(image, question)["answer"]
def _query_qwen(self, image: Image.Image, question: str) -> str:
"""Query Qwen2.5-VL using the chat-template pipeline.
Args:
image: PIL Image.
question: Text question.
Returns:
Generated answer text (input tokens stripped).
"""
messages = [
{
"role": "user",
"content": [
{"type": "image", "image": image},
{"type": "text", "text": question},
],
}
]
# tokenize=False so we get a plain string; processor does tokenisation.
text: str = self.processor.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
# Use process_vision_info from qwen_vl_utils when available β€” it
# correctly handles the image/video extraction from the message dict.
if _has_qwen_vl_utils():
from qwen_vl_utils import process_vision_info # type: ignore[import]
image_inputs, video_inputs = process_vision_info(messages)
else:
image_inputs = [image]
video_inputs = None
processor_kwargs: dict = dict(
text=[text],
images=image_inputs,
padding=True,
return_tensors="pt",
)
if video_inputs is not None:
processor_kwargs["videos"] = video_inputs
inputs = self.processor(**processor_kwargs).to(
"cuda" if torch.cuda.is_available() else "cpu"
)
with torch.inference_mode():
generated_ids = self.model.generate(
**inputs,
max_new_tokens=256,
do_sample=False,
)
# Trim the input tokens so batch_decode only sees new tokens.
trimmed = [
out[len(inp):]
for inp, out in zip(inputs.input_ids, generated_ids)
]
return self.processor.batch_decode(
trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)[0]