|
| 1 | +# Porting Spec-Bench Inference Runners to specdec_bench |
| 2 | + |
| 3 | +This guide explains how to convert any `inference_*.py` runner from [Spec-Bench](https://github.com/hemingkx/Spec-Bench) to a model class compatible with `specdec_bench`. |
| 4 | + |
| 5 | +## Overview |
| 6 | + |
| 7 | +Spec-Bench inference runners follow a pattern where: |
| 8 | +1. A `*_forward()` function handles the speculative decoding logic |
| 9 | +2. The `run_eval()` function orchestrates evaluation with tokenized inputs |
| 10 | +3. Models are loaded in `__main__` and passed to `run_eval()` |
| 11 | + |
| 12 | +In contrast, `specdec_bench` uses a class-based approach where: |
| 13 | +1. Models inherit from the `Model` base class |
| 14 | +2. `__init__()` handles model loading |
| 15 | +3. `run()` is an async method that processes single requests |
| 16 | +4. `stop()` handles cleanup |
| 17 | + |
| 18 | +## The specdec_bench Model Interface |
| 19 | + |
| 20 | +```python |
| 21 | +class Model: |
| 22 | + def __init__(self, model_dir, tokenizer, max_draft_length): |
| 23 | + raise NotImplementedError |
| 24 | + |
| 25 | + async def run(self, prompt_ids, sampling_params, request_id, turn_id): |
| 26 | + """ |
| 27 | + prompt_ids: list of token IDs (not a tensor!) |
| 28 | + Returns dict with: |
| 29 | + - output_ids: list of list of token chunks per step [[chunk1, chunk2, ...]] |
| 30 | + - output_logits: optional logits (usually None) |
| 31 | + - token_times: list of timestamps per decoding step |
| 32 | + """ |
| 33 | + raise NotImplementedError |
| 34 | + |
| 35 | + def stop(self): |
| 36 | + pass |
| 37 | +``` |
| 38 | + |
| 39 | +## Step-by-Step Porting Guide |
| 40 | + |
| 41 | +### Step 1: Identify the Key Components in Spec-Bench |
| 42 | + |
| 43 | +Look at the `inference_*.py` file and identify: |
| 44 | + |
| 45 | +1. **The forward function** (e.g., `medusa_forward`, `ea_forward`) |
| 46 | + - This contains the core speculative decoding loop |
| 47 | + - Signature: `forward_func(inputs, model, tokenizer, max_new_tokens, **kwargs)` |
| 48 | + - Returns: `(output_ids, new_token_count, num_steps, accept_length_list)` |
| 49 | + |
| 50 | +2. **The model class** (e.g., `MedusaModel`, `EaModel`) |
| 51 | + - Found in `model/<method>/` directory |
| 52 | + - Has a `from_pretrained()` class method |
| 53 | + |
| 54 | +3. **Required utilities** from the method's module: |
| 55 | + - Buffer generation (e.g., `generate_medusa_buffers`) |
| 56 | + - Initialization functions (e.g., `initialize_medusa`, `initialize_past_key_values`) |
| 57 | + - Decoding functions (e.g., `tree_decoding`, `generate_candidates`) |
| 58 | + - State update functions (e.g., `update_inference_inputs`) |
| 59 | + |
| 60 | +4. **Method-specific choices/configs** (e.g., `mc_sim_7b_63` for Medusa) |
| 61 | + |
| 62 | +### Step 2: Create the specdec_bench Model Class |
| 63 | + |
| 64 | +```python |
| 65 | +# specdec_bench/specdec_bench/models/specbench_<method>.py |
| 66 | + |
| 67 | +from .base import Model |
| 68 | +import asyncio |
| 69 | +import time |
| 70 | +import torch |
| 71 | + |
| 72 | +# Import dependencies from Spec-Bench |
| 73 | +try: |
| 74 | + import sys |
| 75 | + import os |
| 76 | + spec_bench_path = os.path.join(os.getcwd(), "Spec-Bench") |
| 77 | + sys.path.insert(0, spec_bench_path) |
| 78 | + from model.<method>.<model_file> import <ModelClass> |
| 79 | + from model.<method>.kv_cache import initialize_past_key_values |
| 80 | + from model.<method>.utils import ( |
| 81 | + # Import all required utilities |
| 82 | + ) |
| 83 | + from model.<method>.<choices_file> import <default_choices> |
| 84 | +except ImportError as e: |
| 85 | + print(f"<Method> dependencies not found: {e}") |
| 86 | + <ModelClass> = None |
| 87 | + |
| 88 | + |
| 89 | +class SpecBench<Method>Model(Model): |
| 90 | + def __init__(self, model_dir, max_concurrent_requests, sampling_kwargs, **kwargs): |
| 91 | + # 1. Validate dependencies |
| 92 | + if <ModelClass> is None: |
| 93 | + raise ImportError("<Method> dependencies not found.") |
| 94 | + |
| 95 | + # 2. Extract configuration from kwargs |
| 96 | + self.dtype = kwargs.get("dtype", "float16") |
| 97 | + self.max_steps = kwargs.get("max_steps", 512) |
| 98 | + self.temperature = sampling_kwargs.get("temperature", 0.0) |
| 99 | + # ... other method-specific parameters |
| 100 | + |
| 101 | + # 3. Set up device (avoid device_map="auto" for multi-GPU issues) |
| 102 | + self.device = torch.device(kwargs.get("device", "cuda:0")) |
| 103 | + |
| 104 | + # 4. Convert dtype string to torch dtype |
| 105 | + dtype_map = { |
| 106 | + "float32": torch.float32, |
| 107 | + "float16": torch.float16, |
| 108 | + "bfloat16": torch.bfloat16, |
| 109 | + } |
| 110 | + torch_dtype = dtype_map.get(self.dtype, torch.float16) |
| 111 | + |
| 112 | + # 5. Load the model |
| 113 | + self.model = <ModelClass>.from_pretrained( |
| 114 | + model_dir, |
| 115 | + # ... other args from Spec-Bench's __main__ |
| 116 | + torch_dtype=torch_dtype, |
| 117 | + low_cpu_mem_usage=True, |
| 118 | + ) |
| 119 | + self.model = self.model.to(self.device) |
| 120 | + |
| 121 | + self.sampling_kwargs = sampling_kwargs |
| 122 | +``` |
| 123 | + |
| 124 | +### Step 3: Port the Forward Function |
| 125 | + |
| 126 | +Convert the standalone `*_forward()` function to an internal method: |
| 127 | + |
| 128 | +```python |
| 129 | + def _forward(self, input_ids, max_new_tokens, end_id): |
| 130 | + """ |
| 131 | + Port of the original *_forward function. |
| 132 | + |
| 133 | + Key changes from Spec-Bench: |
| 134 | + 1. input_ids is already a tensor (converted in run()) |
| 135 | + 2. Add timing list to track per-step timestamps |
| 136 | + 3. Use self.device instead of model.base_model.device |
| 137 | + 4. Return timing along with other outputs |
| 138 | + """ |
| 139 | + accept_length_list = [] |
| 140 | + timing = [time.perf_counter()] # ADD: Track timing |
| 141 | + |
| 142 | + # === COPY THE FORWARD LOGIC FROM SPEC-BENCH === |
| 143 | + # Replace: device=model.base_model.device |
| 144 | + # With: device=self.device |
| 145 | + |
| 146 | + # Initialize buffers... |
| 147 | + # Initialize KV cache... |
| 148 | + # Main decoding loop... |
| 149 | + |
| 150 | + for idx in range(self.max_steps): |
| 151 | + # Generate candidates... |
| 152 | + # Tree decoding... |
| 153 | + # Evaluate posterior... |
| 154 | + # Update inputs... |
| 155 | + |
| 156 | + timing.append(time.perf_counter()) # ADD: Record time per step |
| 157 | + |
| 158 | + # Check for EOS |
| 159 | + if end_id in input_ids[0, input_len:].tolist(): |
| 160 | + break |
| 161 | + if new_token > max_new_tokens: |
| 162 | + break |
| 163 | + |
| 164 | + return input_ids, new_token, idx + 1, accept_length_list, timing # ADD timing |
| 165 | +``` |
| 166 | + |
| 167 | +### Step 4: Implement the run() Method |
| 168 | + |
| 169 | +```python |
| 170 | + async def run(self, prompt_ids, max_length, end_id, request_id, turn_id): |
| 171 | + """ |
| 172 | + Async interface for specdec_bench. |
| 173 | + |
| 174 | + Args: |
| 175 | + prompt_ids: List of input token IDs (NOT a tensor) |
| 176 | + max_length: Maximum new tokens to generate |
| 177 | + end_id: EOS token ID |
| 178 | + request_id: Request identifier |
| 179 | + turn_id: Turn identifier |
| 180 | + |
| 181 | + Returns: |
| 182 | + dict with output_ids, output_logits, token_times |
| 183 | + """ |
| 184 | + output_dict = {} |
| 185 | + |
| 186 | + # Convert prompt_ids list to tensor |
| 187 | + input_ids = torch.tensor( |
| 188 | + [prompt_ids], dtype=torch.long, device=self.device |
| 189 | + ) |
| 190 | + |
| 191 | + # Run forward pass (use asyncio.to_thread for sync code) |
| 192 | + result = await asyncio.to_thread( |
| 193 | + self._forward, input_ids, max_length, end_id |
| 194 | + ) |
| 195 | + input_ids_out, new_token, num_steps, accept_length_list, timing = result |
| 196 | + |
| 197 | + # Extract generated tokens (excluding prompt) |
| 198 | + original_len = len(prompt_ids) |
| 199 | + generated_tokens = input_ids_out[0, original_len:].tolist() |
| 200 | + |
| 201 | + # Remove EOS token if present |
| 202 | + if end_id in generated_tokens: |
| 203 | + eos_idx = generated_tokens.index(end_id) |
| 204 | + generated_tokens = generated_tokens[:eos_idx] |
| 205 | + |
| 206 | + # Format output_ids as list of token chunks per step |
| 207 | + # This matches specdec_bench's expected format |
| 208 | + reformatted_output_ids = [[]] |
| 209 | + start = 0 |
| 210 | + for accept_len in accept_length_list: |
| 211 | + if accept_len > 0 and start < len(generated_tokens): |
| 212 | + chunk = generated_tokens[start:start + accept_len] |
| 213 | + if chunk: |
| 214 | + reformatted_output_ids[0].append(chunk) |
| 215 | + start += accept_len |
| 216 | + |
| 217 | + # Handle remaining tokens |
| 218 | + if start < len(generated_tokens): |
| 219 | + reformatted_output_ids[0].append(generated_tokens[start:]) |
| 220 | + |
| 221 | + output_dict['output_ids'] = reformatted_output_ids |
| 222 | + output_dict['output_logits'] = None |
| 223 | + output_dict['token_times'] = timing |
| 224 | + |
| 225 | + return output_dict |
| 226 | +``` |
| 227 | + |
| 228 | +### Step 5: Implement stop() for Cleanup |
| 229 | + |
| 230 | +```python |
| 231 | + def stop(self): |
| 232 | + """Clean up resources.""" |
| 233 | + # Clear any cached states |
| 234 | + if hasattr(self.model, "past_key_values"): |
| 235 | + del self.model.past_key_values |
| 236 | + del self.model.past_key_values_data |
| 237 | + del self.model.current_length_data |
| 238 | + |
| 239 | + # Clear method-specific buffers |
| 240 | + if hasattr(self.model, "<method>_buffers"): |
| 241 | + del self.model.<method>_buffers |
| 242 | + |
| 243 | + # Free GPU memory |
| 244 | + if hasattr(self, 'model') and self.model is not None: |
| 245 | + del self.model |
| 246 | + torch.cuda.empty_cache() |
| 247 | +``` |
| 248 | + |
| 249 | +### Step 6: Register the Model (Optional) |
| 250 | + |
| 251 | +Add to `specdec_bench/specdec_bench/models/__init__.py`: |
| 252 | + |
| 253 | +```python |
| 254 | +from .specbench_<method> import SpecBench<Method>Model |
| 255 | +``` |
| 256 | + |
| 257 | +## Key Differences Summary |
| 258 | + |
| 259 | +| Aspect | Spec-Bench | specdec_bench | |
| 260 | +|--------|-----------|---------------| |
| 261 | +| Input format | `inputs.input_ids` (tensor from tokenizer) | `prompt_ids` (list of ints) | |
| 262 | +| Output format | `(output_ids, new_token, steps, accept_lengths)` | `dict` with `output_ids`, `output_logits`, `token_times` | |
| 263 | +| Output IDs | Full sequence tensor | List of token chunks per step | |
| 264 | +| Timing | External (in `run_eval`) | Internal (in `run()`) | |
| 265 | +| Device | `device_map="auto"` | Explicit single device | |
| 266 | +| Interface | Function-based | Class-based with async `run()` | |
| 267 | + |
| 268 | +## Common Pitfalls |
| 269 | + |
| 270 | +1. **Device Mismatch**: Avoid `device_map="auto"` which spreads model across GPUs. Use explicit `.to(device)`. |
| 271 | + |
| 272 | +2. **Tensor vs List**: `prompt_ids` in specdec_bench is a Python list, not a tensor. Convert it in `run()`. |
| 273 | + |
| 274 | +3. **Output Format**: specdec_bench expects `output_ids` as `[[chunk1, chunk2, ...]]` (list of lists of lists for beam_width=1). |
| 275 | + |
| 276 | +4. **Timing**: Add `time.perf_counter()` calls to track per-step latency. |
| 277 | + |
| 278 | +5. **EOS Handling**: Strip EOS tokens from output before formatting. |
| 279 | + |
| 280 | +6. **Async Wrapper**: Use `asyncio.to_thread()` to wrap synchronous forward passes. |
| 281 | + |
| 282 | +## Example: Mapping Spec-Bench Methods |
| 283 | + |
| 284 | +| Spec-Bench File | Model Class | Forward Function | Key Utils | |
| 285 | +|-----------------|-------------|------------------|-----------| |
| 286 | +| `inference_medusa.py` | `MedusaModel` | `medusa_forward` | `generate_medusa_buffers`, `initialize_medusa` | |
| 287 | +| `inference_eagle.py` | `EaModel` | `ea_forward` | `generate_tree_buffers`, `initialize_tree` | |
| 288 | +| `inference_eagle2.py` | `EaModel` | `ea_forward` | Same as EAGLE | |
| 289 | +| `inference_hydra.py` | `HydraModel` | `hydra_forward` | `generate_hydra_buffers`, `initialize_hydra` | |
| 290 | +| `inference_lookahead.py` | `LookaheadModel` | `lookahead_forward` | Lookahead-specific utils | |
| 291 | + |
| 292 | +## Testing Your Port |
| 293 | + |
| 294 | +```python |
| 295 | +import asyncio |
| 296 | + |
| 297 | +async def test(): |
| 298 | + model = SpecBench<Method>Model( |
| 299 | + model_dir="/path/to/model", |
| 300 | + max_concurrent_requests=1, |
| 301 | + sampling_kwargs={"temperature": 0.0}, |
| 302 | + # method-specific kwargs... |
| 303 | + ) |
| 304 | + |
| 305 | + result = await model.run( |
| 306 | + prompt_ids=[1, 2, 3, 4, 5], # Example token IDs |
| 307 | + max_length=100, |
| 308 | + end_id=2, # EOS token |
| 309 | + request_id="test", |
| 310 | + turn_id=0 |
| 311 | + ) |
| 312 | + |
| 313 | + print("Output chunks:", result['output_ids']) |
| 314 | + print("Timing:", result['token_times']) |
| 315 | + |
| 316 | + model.stop() |
| 317 | + |
| 318 | +asyncio.run(test()) |
| 319 | +``` |
0 commit comments