handler.py

"""Custom inference handler for HuggingFace Inference Endpoints."""

from typing import Any, Dict, List, Union

import torch

try:
    # For remote execution, imports are relative
    from .asr_modeling import ASRModel
    from .asr_pipeline import ASRPipeline
except ImportError:
    # For local execution, imports are not relative
    from asr_modeling import ASRModel  # type: ignore[no-redef]
    from asr_pipeline import ASRPipeline  # type: ignore[no-redef]


class EndpointHandler:
    def __init__(self, path: str = ""):
        # Set environment variables for PyTorch/CUDA (must be before imports/operations)
        import os

        # Enable expandable segments to reduce fragmentation
        os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")

        # Enable TF32 for faster matmul on A40/A100
        torch.backends.cuda.matmul.allow_tf32 = True
        torch.backends.cudnn.allow_tf32 = True

        # Set device and dtype
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.dtype = torch.bfloat16 if self.device == "cuda" else torch.float32

        # Enable CUDA optimizations
        if torch.cuda.is_available():
            torch.backends.cudnn.benchmark = True

        # Prepare model kwargs for pipeline
        model_kwargs = {
            "dtype": self.dtype,
            "low_cpu_mem_usage": True,
        }
        if torch.cuda.is_available():
            model_kwargs["attn_implementation"] = (
                "flash_attention_2" if self._is_flash_attn_available() else "sdpa"
            )

        # Load model (this loads the model, tokenizer, and feature extractor)
        self.model = ASRModel.from_pretrained(path, **model_kwargs)

        # Instantiate custom pipeline - it will get feature_extractor and tokenizer from model
        self.pipe = ASRPipeline(
            model=self.model,
            feature_extractor=self.model.feature_extractor,
            tokenizer=self.model.tokenizer,
            device=self.device,
        )

        # Apply torch.compile if enabled (after model is loaded by pipeline)
        # Enable by default for significant speedup (20-40%)
        if torch.cuda.is_available() and os.getenv("ENABLE_TORCH_COMPILE", "1") == "1":
            compile_mode = os.getenv("TORCH_COMPILE_MODE", "reduce-overhead")
            self.model = torch.compile(self.model, mode=compile_mode)
            # Update the pipeline with the compiled model
            self.pipe.model = self.model

        # Warmup the model
        if torch.cuda.is_available():
            self._warmup()

    def _is_flash_attn_available(self):
        """Check if flash attention is available."""
        import importlib.util

        return importlib.util.find_spec("flash_attn") is not None

    def _warmup(self):
        """Warmup to trigger model compilation and allocate GPU memory."""
        try:
            # Create dummy audio (1 second at config sample rate)
            sample_rate = self.pipe.model.config.audio_sample_rate
            dummy_audio = torch.randn(sample_rate, dtype=torch.float32)

            # The pipeline now handles GPU optimization internally
            with torch.inference_mode():
                warmup_tokens = self.pipe.model.config.inference_warmup_tokens
                _ = self.pipe(
                    {"raw": dummy_audio, "sampling_rate": sample_rate}, max_new_tokens=warmup_tokens
                )

            # Force CUDA synchronization to ensure kernels are compiled
            if torch.cuda.is_available():
                torch.cuda.synchronize()
                # Clear cache after warmup to free memory
                torch.cuda.empty_cache()

        except Exception as e:
            print(f"Warmup skipped due to: {e}")

    def __call__(self, data: Dict[str, Any]) -> Union[Dict[str, Any], List[Dict[str, Any]]]:
        inputs = data.get("inputs")
        if inputs is None:
            raise ValueError("Missing 'inputs' in request data")

        params = data.get("parameters", {})
        max_new_tokens = params.get("max_new_tokens", 200)
        num_beams = params.get("num_beams", 1)
        do_sample = params.get("do_sample", False)
        length_penalty = params.get("length_penalty", 1.0)
        repetition_penalty = params.get("repetition_penalty", 1.0)
        no_repeat_ngram_size = params.get("no_repeat_ngram_size", 0)
        early_stopping = params.get("early_stopping", True)
        default_diversity = self.pipe.model.config.inference_diversity_penalty
        diversity_penalty = params.get("diversity_penalty", default_diversity)

        return self.pipe(
            inputs,
            max_new_tokens=max_new_tokens,
            num_beams=num_beams,
            do_sample=do_sample,
            length_penalty=length_penalty,
            repetition_penalty=repetition_penalty,
            no_repeat_ngram_size=no_repeat_ngram_size,
            early_stopping=early_stopping,
            diversity_penalty=diversity_penalty,
        )