Update custom model files, README, and requirements

README.md

@@ -1,199 +1,73 @@
 ---
-library_name: transformers
-tags: []
+license: mit
+language:
+- en
+datasets:
+- speechbrain/LoquaciousSet
+base_model:
+- openai/whisper-large-v3-turbo
+- HuggingFaceTB/SmolLM3-3B
+pipeline_tag: automatic-speech-recognition
+tags:
+- asr
+- speech-recognition
+- audio
+- smollm
+- whisper
+- mlp
 ---
 
-# Model Card for Model ID
+# Tiny Audio
 
-<!-- Provide a quick summary of what the model is/does. -->
+A speech recognition model trained in 24 hours on a single GPU for ~$12. Built with the [Tiny Audio](https://github.com/alexkroman/tiny-audio) codebase—a minimal, hackable framework for training ASR models.
 
+## Architecture
 
+```
+Audio (16kHz) → Whisper Encoder (frozen) → MLP Projector (trained) → SmolLM3-3B (frozen) → Text
+```
 
-## Model Details
-
-### Model Description
-
-<!-- Provide a longer summary of what this model is. -->
-
-This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
-
-- **Developed by:** [More Information Needed]
-- **Funded by [optional]:** [More Information Needed]
-- **Shared by [optional]:** [More Information Needed]
-- **Model type:** [More Information Needed]
-- **Language(s) (NLP):** [More Information Needed]
-- **License:** [More Information Needed]
-- **Finetuned from model [optional]:** [More Information Needed]
-
-### Model Sources [optional]
-
-<!-- Provide the basic links for the model. -->
-
-- **Repository:** [More Information Needed]
-- **Paper [optional]:** [More Information Needed]
-- **Demo [optional]:** [More Information Needed]
-
-## Uses
-
-<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
-
-### Direct Use
-
-<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
-
-[More Information Needed]
-
-### Downstream Use [optional]
-
-<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
-
-[More Information Needed]
-
-### Out-of-Scope Use
-
-<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
-
-[More Information Needed]
-
-## Bias, Risks, and Limitations
-
-<!-- This section is meant to convey both technical and sociotechnical limitations. -->
-
-[More Information Needed]
-
-### Recommendations
-
-<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
-
-Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
-
-## How to Get Started with the Model
-
-Use the code below to get started with the model.
-
-[More Information Needed]
+**MLP Projector:**
+- Convolutional downsampling: 4x sequence compression via two stride-2 conv layers
+- Linear (1280 → 2048) → GELU → Linear (2048 → 2048)
+- Output normalization: RMSNorm
 
 ## Training Details
 
-### Training Data
-
-<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
-
-[More Information Needed]
-
-### Training Procedure
-
-<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
-
-#### Preprocessing [optional]
-
-[More Information Needed]
-
-
-#### Training Hyperparameters
-
-- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
-
-#### Speeds, Sizes, Times [optional]
-
-<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
-
-[More Information Needed]
-
-## Evaluation
-
-<!-- This section describes the evaluation protocols and provides the results. -->
-
-### Testing Data, Factors & Metrics
-
-#### Testing Data
-
-<!-- This should link to a Dataset Card if possible. -->
-
-[More Information Needed]
-
-#### Factors
-
-<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
-
-[More Information Needed]
-
-#### Metrics
-
-<!-- These are the evaluation metrics being used, ideally with a description of why. -->
-
-[More Information Needed]
-
-### Results
-
-[More Information Needed]
-
-#### Summary
-
-
-
-## Model Examination [optional]
-
-<!-- Relevant interpretability work for the model goes here -->
-
-[More Information Needed]
-
-## Environmental Impact
-
-<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
-
-Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
-
-- **Hardware Type:** [More Information Needed]
-- **Hours used:** [More Information Needed]
-- **Cloud Provider:** [More Information Needed]
-- **Compute Region:** [More Information Needed]
-- **Carbon Emitted:** [More Information Needed]
-
-## Technical Specifications [optional]
-
-### Model Architecture and Objective
-
-[More Information Needed]
-
-### Compute Infrastructure
-
-[More Information Needed]
-
-#### Hardware
-
-[More Information Needed]
-
-#### Software
-
-[More Information Needed]
-
-## Citation [optional]
-
-<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
-
-**BibTeX:**
-
-[More Information Needed]
+| | |
+|---|---|
+| **Dataset** | LoquaciousSet (25,000 hours) |
+| **Hardware** | Single NVIDIA A40 40GB |
+| **Training Time** | ~24 hours |
+| **Cost** | ~$12 |
+| **Trainable Parameters** | ~12M (projector only) |
 
-**APA:**
+## Performance
 
-[More Information Needed]
+**Word Error Rate (WER): 12.14%** on LoquaciousSet test set.
 
-## Glossary [optional]
+See the [community leaderboard](https://github.com/alexkroman/tiny-audio#leaderboard) for comparisons.
 
-<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+## Usage
 
-[More Information Needed]
+```python
+from transformers import pipeline
 
-## More Information [optional]
+pipe = pipeline("automatic-speech-recognition", model="mazesmazes/tiny-audio", trust_remote_code=True)
 
-[More Information Needed]
+result = pipe("path/to/audio.wav")
+print(result["text"])
+```
 
-## Model Card Authors [optional]
+## Limitations
 
-[More Information Needed]
+- English only
+- Optimized for 16kHz audio; other sample rates are resampled automatically
+- Performance may degrade on heavily accented speech, noisy environments, or domain-specific jargon
+- Maximum audio length limited by context window
 
-## Model Card Contact
+## Learn More
 
-[More Information Needed]
+- **[Train your own model](https://github.com/alexkroman/tiny-audio)** — The full codebase with training scripts
+- **[Free 3-hour course](https://github.com/alexkroman/tiny-audio/blob/main/docs/course/0-course-overview.md)** — Build your own ASR system from scratch
+- **[Submit to leaderboard](https://github.com/alexkroman/tiny-audio#leaderboard)** — Share your trained model

asr_config.py

@@ -18,6 +18,7 @@ class ASRConfig(transformers.PretrainedConfig):
         user_prompt: str = "Transcribe: <audio>",
         encoder_dim: Optional[int] = None,
         llm_dim: Optional[int] = None,
+        encoder_stride: int = 2,  # Temporal downsampling factor of audio encoder
         audio_sample_rate: int = 16000,
         projector_init_std: float = 0.02,
         projector_pool_stride: int = 4,
@@ -26,6 +27,7 @@ class ASRConfig(transformers.PretrainedConfig):
         projector_type: str = "moe",  # "moe", "swiglu", "residual", "shared_moe", "mlp", "qformer"
         projector_num_layers: int = 2,  # Number of layers (for residual projector)
         projector_dropout: float = 0.0,  # Dropout rate for projector layers
+        projector_downsample: bool = True,  # Whether to downsample in MLP projector
         # MoE-specific configuration
         num_experts: int = 4,  # Number of experts in MoE projectors
         num_experts_per_tok: int = 2,  # Top-k experts per token
@@ -66,6 +68,7 @@ class ASRConfig(transformers.PretrainedConfig):
         self.user_prompt = user_prompt
         self.encoder_dim = encoder_dim
         self.llm_dim = llm_dim
+        self.encoder_stride = encoder_stride
         self.audio_sample_rate = audio_sample_rate
         self.projector_init_std = projector_init_std
         self.projector_pool_stride = projector_pool_stride
@@ -74,6 +77,7 @@ class ASRConfig(transformers.PretrainedConfig):
         self.projector_type = projector_type
         self.projector_num_layers = projector_num_layers
         self.projector_dropout = projector_dropout
+        self.projector_downsample = projector_downsample
         # MoE-specific configuration
         self.num_experts = num_experts
         self.num_experts_per_tok = num_experts_per_tok

asr_modeling.py

@@ -81,6 +81,7 @@ class ASRModel(PreTrainedModel, GenerationMixin):
         super().__init__(config)
 
         self.system_prompt = config.system_prompt
+        self.encoder_stride = config.encoder_stride
         target_dtype = getattr(torch, config.model_dtype)
 
         # Audio encoder (frozen)
@@ -138,6 +139,17 @@ class ASRModel(PreTrainedModel, GenerationMixin):
             full_model = WhisperModel.from_pretrained(config.audio_model_id, **encoder_kwargs)
             encoder = full_model.encoder
             del full_model
+        elif "glm" in config.audio_model_id.lower():
+            # GLM-ASR models use audio_tower as the encoder
+            # Requires transformers >= 5.x or installed from source
+            from transformers import AutoModelForSeq2SeqLM
+
+            full_model = AutoModelForSeq2SeqLM.from_pretrained(
+                config.audio_model_id, trust_remote_code=True, **encoder_kwargs
+            )
+            # GLM stores encoder at audio_tower (GlmAsrEncoder)
+            encoder = full_model.audio_tower
+            del full_model
         else:
             encoder = AutoModel.from_pretrained(config.audio_model_id, **encoder_kwargs)
 
@@ -207,7 +219,7 @@ class ASRModel(PreTrainedModel, GenerationMixin):
             self.tokenizer.pad_token = "<|finetune_right_pad_id|>"
 
         # Add audio token
-        existing_special = self.tokenizer.additional_special_tokens or []
+        existing_special = getattr(self.tokenizer, "additional_special_tokens", None) or []
         if "<audio>" not in existing_special:
             self.tokenizer.add_special_tokens(
                 {"additional_special_tokens": existing_special + ["<audio>"]}
@@ -260,7 +272,12 @@ class ASRModel(PreTrainedModel, GenerationMixin):
         except ImportError:
             from asr_processing import ASRProcessor  # type: ignore[no-redef]
 
-        return ASRProcessor(feature_extractor=self.feature_extractor, tokenizer=self.tokenizer)
+        return ASRProcessor(
+            feature_extractor=self.feature_extractor,
+            tokenizer=self.tokenizer,
+            projector=self.projector,
+            encoder_stride=self.encoder_stride,
+        )
 
     def state_dict(self, *args, **kwargs):
         """Only save trainable projector weights."""
@@ -284,8 +301,8 @@ class ASRModel(PreTrainedModel, GenerationMixin):
             encoder_out = self.audio_tower(input_features=audio_features)
             hidden_states = encoder_out.last_hidden_state
 
-        # Truncate to actual audio length (mel_frames -> encoder_frames via stride-2 conv)
-        real_encoder_len = audio_attention_mask.sum(dim=-1) // 2
+        # Truncate to actual audio length (mel_frames -> encoder_frames via encoder stride)
+        real_encoder_len = audio_attention_mask.sum(dim=-1) // self.encoder_stride
         max_real_len = int(real_encoder_len.max().item())
         hidden_states = hidden_states[:, :max_real_len]
 
@@ -365,10 +382,10 @@ class ASRModel(PreTrainedModel, GenerationMixin):
         """Calculate number of audio tokens based on actual audio length.
 
         Uses attention mask to get real audio length, then computes:
-        mel_frames -> encoder_frames (stride-2) -> projector output tokens
+        mel_frames -> encoder_frames (via encoder stride) -> projector output tokens
         """
         mel_len = int(audio_attention_mask.sum(dim=-1).max().item())
-        encoder_output_len = mel_len // 2
+        encoder_output_len = mel_len // self.encoder_stride
         return int(self.projector.get_output_length(encoder_output_len))
 
     @torch.no_grad()

asr_processing.py

@@ -19,11 +19,12 @@ class ASRProcessor(ProcessorMixin):
     AUDIO_TOKEN = "<audio>"
     TRANSCRIBE_PROMPT = "Transcribe: "
 
-    def __init__(self, feature_extractor, tokenizer, projector=None):
+    def __init__(self, feature_extractor, tokenizer, projector=None, encoder_stride: int = 2):
         self.feature_extractor = feature_extractor
         self.tokenizer = tokenizer
         self.audio_token_id = tokenizer.convert_tokens_to_ids(self.AUDIO_TOKEN)
         self.projector = projector
+        self.encoder_stride = encoder_stride
 
     def __call__(
         self,
@@ -60,7 +61,7 @@ class ASRProcessor(ProcessorMixin):
 
             # Use actual audio length (from attention mask) for token count
             real_mel_len = audio_inputs["attention_mask"].sum(dim=-1).max().item()
-            encoder_output_len = real_mel_len // 2
+            encoder_output_len = real_mel_len // self.encoder_stride
             num_audio_tokens = self.projector.get_output_length(encoder_output_len)
         else:
             num_audio_tokens = 0

handler.py

@@ -0,0 +1,114 @@
+"""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 = ""):
+        import os
+
+        import nltk
+
+        nltk.download("punkt_tab", quiet=True)
+
+        os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
+
+        # Enable TF32 for faster matmul on Ampere+ GPUs (A100, etc.)
+        # Also beneficial for T4 (Turing) which supports TensorFloat-32
+        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"
+
+        # Use float16 for better T4 compatibility (bfloat16 not well supported on T4)
+        # T4 has excellent float16 performance with tensor cores
+        self.dtype = torch.float16 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)
+        # Use "default" mode for T4 - better compatibility than "reduce-overhead"
+        # "reduce-overhead" is better for A100+ but can be slower on older GPUs
+        if torch.cuda.is_available() and os.getenv("ENABLE_TORCH_COMPILE", "1") == "1":
+            compile_mode = os.getenv("TORCH_COMPILE_MODE", "default")
+            self.model = torch.compile(self.model, mode=compile_mode)
+            self.pipe.model = self.model
+
+        # Warmup the model to trigger compilation and optimize kernels
+        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)
+
+            # Run inference to trigger torch.compile and kernel optimization
+            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")
+
+        # Pass through any parameters from request, let model config provide defaults
+        params = data.get("parameters", {})
+
+        return self.pipe(inputs, **params)

projectors.py

@@ -23,45 +23,48 @@ from transformers.models.llama.modeling_llama import LlamaRMSNorm
 
 
 class MLPAudioProjector(nn.Module):
-    """2-layer MLP projector with conv-based 2x temporal downsampling."""
+    """2-layer MLP projector with optional conv-based 2x temporal downsampling.
+
+    When projector_downsample=True (default): Uses Conv1d stride-2 for 2x downsampling.
+    When projector_downsample=False: No downsampling, just MLP projection (like GLM-ASR).
+    """
 
     def __init__(self, config):
         super().__init__()
 
         encoder_dim = getattr(config, "encoder_dim", 768)
         llm_dim = getattr(config, "llm_dim", 2048)
+        self.use_downsample = getattr(config, "projector_downsample", True)
+
+        if self.use_downsample:
+            self.downsample = nn.Conv1d(
+                encoder_dim, encoder_dim, kernel_size=3, stride=2, padding=1, bias=False
+            )
+        else:
+            self.downsample = None
 
-        self.downsample = nn.Conv1d(
-            encoder_dim, encoder_dim, kernel_size=3, stride=2, padding=1, bias=False
-        )
         self.linear_1 = nn.Linear(encoder_dim, llm_dim, bias=False)
         self.act = nn.GELU()
         self.linear_2 = nn.Linear(llm_dim, llm_dim, bias=False)
-
-        self.apply(self._init_weights)
-
-    def _init_weights(self, module):
-        if isinstance(module, nn.Linear):
-            nn.init.normal_(module.weight, mean=0.0, std=0.02)
-        elif isinstance(module, nn.Conv1d):
-            nn.init.normal_(module.weight, mean=0.0, std=0.02)
-            if module.bias is not None:
-                nn.init.zeros_(module.bias)
+        # Using PyTorch default initialization (like GLM-ASR)
 
     def get_output_length(self, input_length: int) -> int:
         """Calculate output sequence length given input length."""
-        # Conv stride=2 halves the length (with padding=1, kernel=3)
-        return (input_length + 1) // 2
+        if self.use_downsample:
+            # Conv stride=2 halves the length (with padding=1, kernel=3)
+            return (input_length + 1) // 2
+        return input_length
 
     def forward(self, x):
         """
         x: [Batch, Seq_Len, Dim]
-        Returns: [Batch, Seq_Len // 2, llm_dim]
+        Returns: [Batch, Seq_Len // 2, llm_dim] if downsampling, else [Batch, Seq_Len, llm_dim]
         """
-        # Conv1d expects [Batch, Channels, Seq_Len]
-        x = x.transpose(1, 2)
-        x = self.downsample(x)
-        x = x.transpose(1, 2)
+        if self.downsample is not None:
+            # Conv1d expects [Batch, Channels, Seq_Len]
+            x = x.transpose(1, 2)
+            x = self.downsample(x)
+            x = x.transpose(1, 2)
 
         x = self.linear_1(x)
         x = self.act(x)
@@ -578,84 +581,6 @@ class QFormerAudioProjector(nn.Module):
         return self.linear(query_proj)
 
 
-# =============================================================================
-# Transformer Projector
-# =============================================================================
-
-
-class TransformerAudioProjector(nn.Module):
-    """
-    Transformer Projector (FunASR Style).
-    Projects to LLM dim first, then applies transformer blocks for context mixing.
-    """
-
-    def __init__(self, config):
-        super().__init__()
-        # Default stride 6: Whisper (2x) * Projector (6x) = 12x total → ~8 Hz
-        # Matches FunASR's total stride (6x encoder * 2x projector = 12x)
-        self.k = getattr(config, "projector_pool_stride", 6)
-
-        encoder_dim = config.encoder_dim
-        llm_dim = config.llm_dim
-
-        # Input: Stacked frames (e.g. 1280 * 2 = 2560)
-        in_dim = encoder_dim * self.k
-
-        # FFN hidden dim for initial projection (balanced compression)
-        # 7680 → 4096 → 2048 distributes compression evenly (~2x each layer)
-        ffn_dim = getattr(config, "projector_hidden_dim", None) or 4096
-
-        # FunASR-style projection: linear1 -> relu -> linear2
-        self.linear1 = nn.Linear(in_dim, ffn_dim)
-        self.relu = nn.ReLU()
-        self.linear2 = nn.Linear(ffn_dim, llm_dim)
-
-        # Transformer blocks operating at llm_dim
-        num_layers = getattr(config, "projector_num_layers", 2)
-        if num_layers > 0:
-            encoder_layer = nn.TransformerEncoderLayer(
-                d_model=llm_dim,
-                nhead=getattr(config, "projector_num_heads", 8),
-                dim_feedforward=1024,  # Match FunASR (audio complexity is LLM-independent)
-                dropout=0.0,
-                activation="relu",
-                batch_first=True,
-                norm_first=True,
-            )
-            self.blocks = nn.TransformerEncoder(
-                encoder_layer, num_layers=num_layers, enable_nested_tensor=False
-            )
-        else:
-            self.blocks = None
-
-    def forward(self, x):
-        # x: [Batch, Seq, Dim]
-        batch, seq, dim = x.shape
-
-        # Padding to multiple of k
-        chunk_num = (seq - 1) // self.k + 1
-        pad_num = chunk_num * self.k - seq
-        if pad_num > 0:
-            x = F.pad(x, (0, 0, 0, pad_num))
-
-        # Frame stacking: [B, S, D] -> [B, S/k, D*k]
-        x = x.contiguous().view(batch, chunk_num, dim * self.k)
-
-        # FunASR-style projection to LLM dim
-        x = self.linear1(x)
-        x = self.relu(x)
-        x = self.linear2(x)
-
-        # Transformer context mixing
-        if self.blocks is not None:
-            x = self.blocks(x)
-
-        return x
-
-    def get_output_length(self, input_length: int) -> int:
-        return (input_length - 1) // self.k + 1
-
-
 # =============================================================================
 # Projector Registry
 # =============================================================================
@@ -666,5 +591,4 @@ PROJECTOR_CLASSES = {
     "swiglu": SwiGLUAudioProjector,
     "shared_moe": SharedMoEAudioProjector,
     "qformer": QFormerAudioProjector,
-    "transformer": TransformerAudioProjector,
 }

requirements.txt

@@ -0,0 +1,6 @@
+# Core dependencies for tiny-audio model inference
+# This file is pushed to HuggingFace for model repository
+
+# Transformers - main library for model loading and inference
+transformers>=4.57.0
+truecase