import ast import contextlib import copy import inspect import itertools import json import os import time from abc import ABC, abstractmethod from dataclasses import dataclass, field from functools import lru_cache from pathlib import Path from typing import Any, Callable, Dict, List, Optional, Set, Tuple, Union import torch from cuda.bindings import driver import tensorrt_llm from tensorrt_llm.bindings.internal.runtime import delay_kernel from tensorrt_llm.logger import logger @dataclass(slots=True, unsafe_hash=True) class DynamicTensorSpec: """ A specification for a dynamic tensor dimension. Args: input_idx: The index of the input tensor. dim_idx: The index of the dimension to tune. gen_tuning_buckets: A tuple of values to try or a function generating values. map_to_tuning_buckets: A function to map dimensions to valid values during inference. """ input_idx: int dim_idx: int gen_tuning_buckets: Union[Tuple[int], Callable] = () map_to_tuning_buckets: Callable = lambda x: x @dataclass(slots=True, unsafe_hash=True) class ConstraintSpec: """ A specification for a constraint on a tensor dimension. Args: input_idx: The index of the input tensor. dim_idx: The index of the dimension to constrain. infer_shape: A function to infer the shape of the dimension. """ input_idx: int dim_idx: int infer_shape: Callable @dataclass(kw_only=True) class TuningConfig: """Configuration for autotuning. This class specifies all the tuning configurations for a single tuning process. Args: dynamic_tensor_specs (Tuple[DynamicTensorSpec]): Specifications for how different tensor dimensions should be tuned to optimize performance. Each spec defines: - Which input tensor dimension is dynamic - How to generate tuning values - How to map dimensions to valid values during inference Example: >>> config = TuningConfig( ... dynamic_tensor_specs=( ... DynamicTensorSpec( ... input_idx=0, ... dim_idx=1, ... gen_tuning_buckets=(32, 64, 128), ... map_to_tuning_buckets=lambda x: ((x + 31) // 32) * 32 ... ), ... ) ... ) constraint_specs (Tuple[ConstraintSpec]): Specifications for constraints on tensor dimensions. Each spec defines: - Which input tensor dimension is constrained - How to infer the shape of the dimension based on other dimensions Example: >>> config = TuningConfig( ... constraint_specs=( ... ConstraintSpec( ... input_idx=1, ... dim_idx=2, ... infer_shape=lambda shapes: shapes[0][0] * 2 ... ), ... ) ... ) tune_max_num_tokens (int): The maximum saturation number of tokens to be tuned. During the inference, the input tensor will be saturated with the same value. Or if any value is provided to the choose_one function, the input tensor will be saturated with the provided value. If not provided, the autotuner will not consider the max num tokens. inputs_pre_hook (Callable): A function that takes a list of input tensors, returns a list of modified input tensors. It is called before the input tensors are prepared for the tuning process to match the real data distribution. use_cold_l2_cache (bool): Whether to use cold L2 cache. This flag is to create circular buffer of input tensors to avoid L2 cache hits to simulate cold L2 cache. Notice that not all tuning processes can benefit from this feature. use_cuda_graph (bool): Whether to use CUDA graph for the tuning process. """ dynamic_tensor_specs: Tuple[DynamicTensorSpec, ...] = () constraint_specs: Tuple[ConstraintSpec, ...] = () tune_max_num_tokens: int = None inputs_pre_hook: Callable = None use_cold_l2_cache: bool = False use_cuda_graph: bool = True @dataclass(unsafe_hash=True) class StaticDim: val: int def _opt(self): return self.val @dataclass(unsafe_hash=True) class DynamicDim: '''Range of one dimension''' min: int opt: int max: int def _opt(self): return self.opt Dim = Union[DynamicDim, StaticDim] @dataclass class OptimizationProfile: '''Ranges of all tensors, all dimension ''' shapes: List[List[Dim]] = field(default_factory=lambda: [[]]) def get_hash_key(self): return self.get_opt_shapes() def get_opt_shapes(self): '''Only the opt shapes are considered as hash key ''' # TODO: remove duplicate shape generation opt_shapes = [] for t in self.shapes: opt_shapes.append(tuple([d._opt() for d in t])) return tuple(opt_shapes) #TODO: can/shall we use the torch builtin FakeTensor class? @dataclass class FakeTensor: dtype: torch.dtype device: torch.device shape: List[Dim] class TunableRunner(ABC): @abstractmethod def get_valid_tactics(self, inputs: List[torch.Tensor], profile: OptimizationProfile, **kwargs) -> List[Any]: """One tactic corresponding to one cuda kernel normally, but how to interpret the meaning of tactic is pure internal details of the runner. The autotuner will just pass the tactic value to the forward w/o. any knowledge on what the tactic means. User can choose to implement their own types of tactic for flexibility, such as using a dict-typed to represent a collection of named configs. tactic==-1 has special meaning, means the fallback kernel which should be able to implement any shapes This fallback tactic is needed for 2 reasons: * when the autotuner cannot find a valid tactic in it's cache. * in eager mode, w/o autotunning the custom op should have at least one kernel, which makes the autotuning process an optional process, such that user can opt out. We choose not to have a standalone can_implement function, the tactics returned by get_valid_tactics should return valid kernel for these given input tensors. """ return [-1] def __call__(self, inputs, **kwargs): return self.forward(inputs, **kwargs) @abstractmethod def forward( self, /, # tensors are position only inputs: List[torch.Tensor], *, # all others are keyword args only tactic: Any = -1, do_preparation: bool = False, **kwargs) -> Any: """Forward pass for tunable runners. Args: inputs: List of input tensors (position-only argument) tactic: A arbitrary type that represents a specific kernel config. For instance, it can be an integer number that specifies the unique ID of the implementation tactic to use. -1 (default) represents the fallback tactic that must be implemented to handle any input shapes when autotuning is disabled. do_preparation: When True, allows one-time setup operations to be performed before tactic evaluation begins. These operations are excluded from the performance measurements during autotuning. Notice that anything prepared in this phase should be persistent in the forward and can be accessed by the following forward calls. Returns: Any: Output of the forward pass. """ raise NotImplementedError def unique_id(self): """ Returns a tuple of the unique id of the runner. The unique id will be converted to a string for the cache key. A common practice is to return a tuple of the runner's attributes, for example: return (self.output_dtype, self.attribute_1, ...) Returns: Any: The unique id of the runner, which can be converted to a string for the cache key. """ return tuple(self.__dict__.values()) @contextlib.contextmanager def autotune(tune_mode: bool = True, cache_path: str = None, rank: int = 0): # if cache_path is provided, use the rank-specific file tune_required = tune_mode if cache_path is not None: # check if the rank-specific file exists cache_path_no_ext = os.path.splitext(cache_path)[0] cache_path_no_ext_rank = cache_path_no_ext + f".rank{rank}.json" # if the rank-specific file exists, load it file_exists = os.path.exists(cache_path_no_ext_rank) # if the rank-specific file exists, do not enable tuning mode if file_exists: logger.info( f"[Autotuner] Loading cache from {cache_path_no_ext_rank}") AutoTuner.get().profiling_cache.load_cache(cache_path_no_ext_rank) # record the old tuning mode old_mode = AutoTuner.get().is_tuning_mode AutoTuner.get().is_tuning_mode = tune_required autotune_enabled = tune_required and not old_mode if autotune_enabled: logger.info("[Autotuner] Autotuning process starts ...") try: yield finally: AutoTuner.get().is_tuning_mode = old_mode if autotune_enabled: logger.info("[Autotuner] Autotuning process ends") # save cache if cache_path is not None: logger.info(f"[Autotuner] Saving cache to {cache_path_no_ext_rank}") AutoTuner.get().profiling_cache.save_cache(cache_path_no_ext_rank) @dataclass class AutoTunerStatistics: """Statistics collected by the AutoTuner. Attributes: cache_misses (int): Number of cache misses requiring fallback cache_miss_config_collection (Dict[str, Set[OptimizationProfile]]): Collection of configs that caused cache misses failed_profiling_count (Dict[str, int]): Number of failed profiling attempts per operation tuned_op_profiled_configs (Dict[str, int]): Profiled configurations per operation tuned_op_time_cost (Dict[str, float]): Time cost per operation """ cache_misses: int = 0 cache_miss_config_collection: Dict[str, Set[tuple]] = field(default_factory=dict) failed_profiling_count: Dict[str, Set[Tuple[str, TunableRunner, OptimizationProfile]]] = field( default_factory=dict) tuned_op_profiled_configs: Dict[str, int] = field(default_factory=dict) tuned_op_time_cost: Dict[str, float] = field(default_factory=dict) def __str__(self) -> str: """Return a string representation of collected statistics. """ stats_str = "" stats_str += f"Cache misses: {self.cache_misses}\n" if self.cache_miss_config_collection: stats_str += "Cache miss config collection:\n" for op, profiles in sorted( self.cache_miss_config_collection.items()): stats_str += f" {op}:\n" for profile in sorted(profiles, key=str): stats_str += f" - Config: {profile}\n" if self.tuned_op_profiled_configs: stats_str += "Tuned operations:\n" for op in sorted(self.tuned_op_profiled_configs.keys()): successful = self.tuned_op_profiled_configs[op] failed = len(self.failed_profiling_count[op]) stats_str += f" {op}:\n" stats_str += f" - Successful configs: {successful}\n" stats_str += f" - Failed profiling count: {failed}\n" if failed > 0: stats_str += f" - Failed profiling combinations:\n" for failed_key in self.failed_profiling_count[op]: stats_str += f" - {failed_key}\n" if self.tuned_op_time_cost: stats_str += "Tuned operations time cost:\n" for op in sorted(self.tuned_op_time_cost.keys()): stats_str += f" {op}: {self.tuned_op_time_cost[op] * 1000:.4f} milliseconds\n" return stats_str class AutoTunerProfilingCache: """AutoTunerCache for caching profiling results. The profiling cache can be serialized to disk for persistence across sessions: - Use save_cache() to save the cache after tuning - Use load_cache() to restore cached results before inference - JSON format provides human-readable output and cross-platform compatibility """ def __init__(self): self.cache = {} # Cache metadata for local storage and validation self.lib_version = tensorrt_llm.__version__ self.creation_timestamp = time.time() # gpu_platform self.device_name = torch.cuda.get_device_name() self.device_capability = torch.cuda.get_device_capability() def __setitem__(self, cache_key: Tuple, value: Tuple) -> None: self.cache[cache_key] = value def __getitem__(self, cache_key: Tuple) -> Tuple: return self.cache[cache_key] def __len__(self) -> int: return len(self.cache) def clear(self) -> None: self.cache.clear() def fallback_entry(self) -> Tuple: # runner_id = 0, tactic = -1 return 0, -1, float('inf') def search_cache( self, custom_op: str, runners: List[TunableRunner], input_shapes: Tuple[torch.Size], tuning_config: TuningConfig, ) -> Tuple[bool, int, int, Dict[str, Any], OptimizationProfile]: """Search for cached profiling results matching the current configuration. Args: custom_op (str): The name of the custom operation to be tuned runners (List[TunableRunner]): List of candidate implementations to profile profile (OptimizationProfile): Optimization profile Returns: A tuple containing: [is_cache_hit, runner_id, tactic, stored_profile] runner_id is the index in the current runners list """ for idx, r in enumerate(runners): if (cache_key := self.get_cache_key(custom_op, r, input_shapes, tuning_config)) in self.cache: # Return the current index in runners list, not the cached runner_id cached_runner_id, tactic, min_time = self.cache[cache_key] return True, idx, tactic, min_time return False, *self.fallback_entry() def get_cache_key( self, custom_op: str, runner: TunableRunner, input_shapes: Tuple[torch.Size], tuning_config: TuningConfig, ) -> Tuple: return ( custom_op, runner.__class__.__name__, str(runner.unique_id()), AutoTuner.get()._find_nearest_profile( input_shapes, tuning_config.dynamic_tensor_specs, tuning_config.constraint_specs, tuning_config.tune_max_num_tokens, ), ) def get_specific_custom_op(self, custom_op: str) -> Dict[Tuple, Tuple]: return {k: v for k, v in self.cache.items() if k[0] == custom_op} def save_cache(self, file_path: Union[str, Path]) -> None: """Save the profiling cache to disk in JSON format. Args: file_path: Path where to save the cache Raises: IOError: If file cannot be written Note: The cache is saved in JSON format which provides human-readable output. Some type information may be lost for complex tactic objects. """ file_path = Path(file_path) file_path.parent.mkdir(parents=True, exist_ok=True) try: serializable_cache = self._serialize_cache_to_json() with open(file_path, 'w') as f: json.dump(serializable_cache, f, indent=2, default=str) logger.info( f"[AutoTuner] Successfully saved cache to {file_path} using JSON format" ) except Exception as e: logger.error(f"[AutoTuner] Failed to save cache with JSON: {e}") raise def load_cache(self, file_path: Union[str, Path]) -> None: """Load the profiling cache from disk in JSON format. Args: file_path: Path to the cache file Raises: FileNotFoundError: If cache file doesn't exist IOError: If file cannot be read Note: Loading will replace the current cache contents. The cache is loaded from JSON format. """ file_path = Path(file_path) if not file_path.exists(): raise FileNotFoundError(f"Cache file not found: {file_path}") try: with open(file_path, 'r') as f: serializable_cache = json.load(f) self.cache = self._deserialize_cache_from_json(serializable_cache) logger.info( f"[AutoTuner] Successfully loaded cache from {file_path} using JSON format" ) except Exception as e: logger.error(f"[AutoTuner] Failed to load cache with JSON: {e}") raise def _serialize_cache_to_json(self) -> Dict[str, Any]: """Convert the profiling cache to a JSON-serializable format. Returns: Dictionary that can be serialized to JSON Note: This method handles the conversion of complex objects to JSON-compatible representations. Some type information may be lost in the conversion. """ serializable_cache = { "metadata": { "lib_version": self.lib_version, "creation_timestamp": self.creation_timestamp, "device_name": self.device_name, "device_capability": self.device_capability, }, "cache_data": {}, } for key, value in self.cache.items(): # Convert tuple key to string for JSON compatibility key_str = str(key) runner_id, tactic, min_time = value serializable_cache["cache_data"][key_str] = { "runner_id": runner_id, "tactic": tactic, "min_time": min_time, } return serializable_cache def _deserialize_cache_from_json( self, serializable_cache: Dict[str, Any]) -> Dict[Tuple, Tuple]: """Convert JSON-serialized cache back to the original format. Args: serializable_cache: Dictionary loaded from JSON Returns: Profiling cache in the original format Note: This attempts to reconstruct the original data structures but may not perfectly preserve all type information, especially for complex tactic objects. """ metadata = serializable_cache["metadata"] self.lib_version = metadata["lib_version"] self.creation_timestamp = metadata["creation_timestamp"] self.device_name = metadata["device_name"] self.device_capability = metadata["device_capability"] cache = {} cache_data = serializable_cache["cache_data"] for key_str, value in cache_data.items(): # Reconstruct the tuple key safely try: key = ast.literal_eval(key_str) # Safer than eval() except (ValueError, SyntaxError): logger.warning( f"[AutoTuner] Could not reconstruct cache key: {key_str}") continue runner_id = value["runner_id"] tactic = value["tactic"] min_time = value["min_time"] cache[key] = (runner_id, tactic, min_time) return cache class AutoTuner: """AutoTuner for optimizing TensorRT LLM operations. This class handles automatic performance tuning of tensor operations by profiling different implementations and caching the best performing configurations. Args: warmup (int): Number of warmup iterations before profiling (default: 3) repeat (int): Number of profiling iterations for averaging (default: 10) stream_delay_micro_secs (int): Delay on CUDA stream before the profiled kernel runs in microseconds (default: 1000) """ _CUDA_GRAPH_DELAY_MICRO_SECS = 100 _instance = None def __init__(self, warmup=2, repeat=10, stream_delay_micro_secs=1000): self.repeat = repeat self.warmup = warmup self.stream_delay_micro_secs = stream_delay_micro_secs self.profiling_cache = AutoTunerProfilingCache() self.is_tuning_mode = False # Add statistics tracking self.stats = AutoTunerStatistics() # Current captured choose_one() contexts self._active_capture: Optional['AutoTuner.TacticsCapture'] = None # Last captured choose_one() contexts self._last_capture: Optional['AutoTuner.TacticsCapture'] = None # Increase log level for AutoTuner associated logger self._log_level_to_info = os.getenv( "TLLM_AUTOTUNER_LOG_LEVEL_DEBUG_TO_INFO", '0') == '1' self._debug_logger = logger.info if self._log_level_to_info else logger.debug @classmethod def get(cls): if cls._instance is None: cls._instance = AutoTuner() return cls._instance class TacticsCapture: """Object returned by capture() that can be iterated to get all tactic combinations. This class encapsulates all state related to capturing and replaying tactics: - Captured execution contexts - Generated tactic configurations - Current replay state (which config and call index) """ def __init__(self, autotuner): # State for captured contexts self._captured_contexts: List[Dict[str, Any]] = [] self._configurations = None # State for replay mode self._replay_runner_tactic_list: Optional[List[Tuple[int, int]]] = None self._replay_context_idx: int = 0 def __iter__(self): """Iterate through all tactic configurations. For single context: yields (runner, tactic) For multiple contexts: yields ((runner_ctx0, tactic_ctx0), (runner_ctx1, tactic_ctx1), ...) """ if self._configurations is None: self._configurations = self._generate_configurations() for config in self._configurations: # config is a tuple of (runner_idx, tactic) for each context # Convert to (runner, tactic) format for user runner_tactic_pairs = [] for ctx_idx, (runner_idx, tactic) in enumerate(config): runners = self._captured_contexts[ctx_idx]['runners'] runner = runners[runner_idx] runner_tactic_pairs.append((runner, tactic)) yield tuple(runner_tactic_pairs) def _generate_configurations(self): """Generate all valid tactic combinations.""" if not self._captured_contexts: raise RuntimeError( "No context available for testing.\n" "Use capture() to capture the operation context first:\n" " with AutoTuner.get().capture() as tactics_capture:\n" " output = operation.forward(...)\n") # Collect valid tactics for each context separately context_tactics_lists = [] for context in self._captured_contexts: runners = context['runners'] inputs = context['inputs'] kwargs = context.get('kwargs', {}) # Collect all valid (runner, tactic) combinations for this context tactics_lists = [] for runner_idx, runner in enumerate(runners): valid_tactics = runner.get_valid_tactics( inputs, OptimizationProfile(), **kwargs) for tactic in valid_tactics: tactics_lists.append((runner_idx, tactic)) context_tactics_lists.append(tactics_lists) # Generate cartesian product from context and tactics where all_configrations[i][ctx] = (runner, tactic) # Such that each element in all_configrations is a replay of multiple contexts of all possible replays all_configurations = list(itertools.product(*context_tactics_lists)) return all_configurations def is_replaying(self) -> bool: """Check if this TacticsCapture is currently in replay mode.""" return self._replay_runner_tactic_list is not None def choose_one( self, custom_op: str, runners: List[TunableRunner], tuning_config: TuningConfig, inputs: List[torch.Tensor], **kwargs, ) -> Tuple: """Choose the best runner and tactic combination through performance profiling. Args: custom_op (str): The name of the custom operation to be tuned runners (List[TunableRunner]): List of candidate implementations to profile tuning_config (TuningConfig): Configuration for the tuning process inputs (List[torch.Tensor]): Input tensors for profiling **kwargs: Arbitrary keyword arguments, will be passed to get_valid_tactics and forward method of each runner Returns: Tuple: A tuple containing: - The selected runner implementation - The best tactic ID for that runner (-1 if using fallback) - The best config for that runner (if configs is not empty) Note: The method profiles different implementations and tactics to find the optimal combination based on performance measurements. It caches results to avoid redundant profiling of the same configuration. Although runners[0] with tactic=-1 is always treated as the fallback runner. Runner authors are suggested to provide a fallback implementation for each runner to avoid potential issues. """ # Check if we're in replay mode via active TacticsCapture if self._active_capture is not None and self._active_capture.is_replaying( ): tactics_capture = self._active_capture call_idx = tactics_capture._replay_context_idx assert call_idx < len(tactics_capture._replay_runner_tactic_list ), "call_idx out of range" assert call_idx < len( tactics_capture._captured_contexts), "call_idx out of range" assert len(tactics_capture._replay_runner_tactic_list) == len( tactics_capture._captured_contexts) # Check if we have a forced tactic for this call and both custom_op match captured_custom_op = tactics_capture._captured_contexts[ call_idx].get('custom_op') if captured_custom_op != custom_op: raise RuntimeError( f"Custom op mismatch in kernel testing mode.\n" f"Expected operation: '{captured_custom_op}'\n" f"Actual operation: '{custom_op}'\n" f"Context index: {call_idx}\n" f"Make sure the forward() call in test mode uses the same operation as captured." ) runner_idx, tactic = tactics_capture._replay_runner_tactic_list[ call_idx] # Increment context counter tactics_capture._replay_context_idx += 1 # Reset counter after all contexts have been used if tactics_capture._replay_context_idx >= len( tactics_capture._replay_runner_tactic_list): tactics_capture._replay_context_idx = 0 return (runners[runner_idx], tactic) # Capture context for testing all underlying kernels if self._active_capture is not None and not self._active_capture.is_replaying( ): self._active_capture._captured_contexts.append({ 'custom_op': custom_op, 'runners': runners, 'tuning_config': tuning_config, 'inputs': inputs, 'kwargs': kwargs, }) input_shapes = tuple(self._get_input_sizes(inputs)) is_cache_hit, best_runner_id, best_tactic, min_time = self.profiling_cache.search_cache( custom_op, runners, input_shapes, tuning_config) # Early return if it's not tuning, use cache found one or fallback one if not self.is_tuning_mode: best_runner = runners[best_runner_id] # TODO: check the stored runner and tactic can implement this shape here # Log the cache miss. Expect no cache miss in inference. if not is_cache_hit: logger.warning_once( f"[AutoTunner] Using the fallback tactic, due to cache miss on input shapes={input_shapes}", key=(custom_op, "warning_autotuning_cache_miss_fallback")) return (best_runner, best_tactic) # If it's tuning mode and cache hit, return the best runner and tactic to avoid redundant profiling. if self.is_tuning_mode and is_cache_hit: return (runners[best_runner_id], best_tactic) assert len(runners) > 0, "At least one runner is required" assert all([isinstance(r, TunableRunner) for r in runners]), \ "All Given runners must be subclass of TunableRunner" tuning_start_time = time.perf_counter() profiles = self._optimization_profiles(tuning_config, inputs) # Initialize the statistics for the custom_op if custom_op not in self.stats.tuned_op_profiled_configs: self.stats.tuned_op_profiled_configs[custom_op] = 0 if custom_op not in self.stats.failed_profiling_count: self.stats.failed_profiling_count[custom_op] = set() new_tuning_failure_occured = False for p in profiles: tensors = self._prepare_input_tensors(p, inputs) is_cache_hit, *_ = self.profiling_cache.search_cache( custom_op, runners, p.get_opt_shapes(), tuning_config) if not is_cache_hit: # Initialize runner and tactic as None in case of no valid tactic or runners are found best_runner_id, best_tactic, min_time, has_tuning_failure_occured = self._profile_runners( custom_op, runners, tensors, p, tuning_config, **kwargs) if best_runner_id is not None: # At least one valid (runner, tactic) pair is found cache_key = self.profiling_cache.get_cache_key( custom_op, runners[best_runner_id], p.get_opt_shapes(), tuning_config) self._debug_logger( f"[Autotuner] Profiling runner={runners[best_runner_id]}, tactic={best_tactic} for cache_key={cache_key}." ) # inspect call stack self.profiling_cache[cache_key] = (best_runner_id, best_tactic, min_time) self.stats.tuned_op_profiled_configs[custom_op] += 1 else: logger.warning_once( f"[Autotuner] No valid runner/tactic was found for custom_op={custom_op}, input_shapes={input_shapes}. " f"At least one valid (runner, tactic) pair is required. " f"If get_valid_tactics is intended to return empty list, please ensure that this profile is not valid for the custom_op " f"and should not occurs during the inference stage, or fallback tactic is implemented. Otherwise, the the tuning process will crash.", key=(custom_op, "warning_autotuning_no_valid_tactic"), ) new_tuning_failure_occured = new_tuning_failure_occured or has_tuning_failure_occured # If failed profiling tactics occurs, log the error. if new_tuning_failure_occured: logger.warning_once( f"[Autotuner] New tuning error occurs:" f"Total failed profiling tactics occurs: {len(self.stats.failed_profiling_count[custom_op])} for custom_op={custom_op}. " f"This will not block the tuning process. " f"Please set TLLM_LOG_LEVEL=WARNING to find out when the tactic profiling fails. " f"Set TLLM_LOG_LEVEL=DEBUG to get more details of the failures.", key=(custom_op, "warning_autotuning_tuning_error_summary"), ) # Get the best runner and tactic from cache # If no valid tactic is found, the fallback runner and tactic will be used _, runner_id, tactic, _ = self.profiling_cache.search_cache( custom_op, runners, input_shapes, tuning_config) tuning_end_time = time.perf_counter() self.stats.tuned_op_time_cost[ custom_op] = self.stats.tuned_op_time_cost.get( custom_op, 0) + tuning_end_time - tuning_start_time return (runners[runner_id], tactic) def _profile_runners( self, custom_op: str, runners: List[TunableRunner], input_tensors: List[torch.Tensor], profile: OptimizationProfile, tuning_config: TuningConfig, **kwargs, ) -> float: min_time = float('inf') has_tuning_failure_occured = False best_runner_id, best_tactic = None, None # If the inputs_pre_hook is provided, it will be called before profiling. if tuning_config.inputs_pre_hook is not None: input_tensors = tuning_config.inputs_pre_hook(input_tensors) for runner_id, runner in enumerate(runners): # TODO: use FakeTensor here. runner_arg_names = { p.name for p in inspect.signature(runner.forward).parameters.values() } valid_tactics = runner.get_valid_tactics(input_tensors, profile, **kwargs) if "do_preparation" in runner_arg_names and len(valid_tactics) > 0: runner( input_tensors, tactic=-1, do_preparation=True, **kwargs, ) for tac in valid_tactics: try: time_measured = self._profile_single_kernel( runner=runner, inputs=input_tensors, tactic=tac, tuning_config=tuning_config, use_cuda_graph=tuning_config.use_cuda_graph, **kwargs, ) except Exception as e: # Handle None tensors for optional inputs shapes = self._get_input_sizes(input_tensors) logger.warning_once( f"[Autotuner] Failed when profiling runner={runner}, tactic={tac}, shapes={shapes}. Set TLLM_LOG_LEVEL=DEBUG for more details.", key=(custom_op, "warning_autotuning_profile_failure"), ) (logger.info_once if self._log_level_to_info else logger.debug_once)( f"[Autotuner] Exception captured: {e}", key=(custom_op, "debug_autotuning_exception"), ) # Record the failed profiling combinations self.stats.failed_profiling_count[custom_op].add( self.profiling_cache.get_cache_key( custom_op, runner, profile.get_opt_shapes(), tuning_config)) # Set time_measured to inf to notify the failure of the tactic. This can happen when `get_valid_tactics` mistakenly return wrong tactics # or some runtime error occurs during profiling. time_measured = float('inf') has_tuning_failure_occured = True if time_measured < min_time: min_time = time_measured best_runner_id, best_tactic = runner_id, tac return best_runner_id, best_tactic, min_time, has_tuning_failure_occured def _get_input_sizes(self, inputs: List[torch.Tensor]) -> List[torch.Size]: # Handle None tensors for optional inputs and non-Tensor scalar values sizes = [ input.size() if isinstance(input, torch.Tensor) else torch.Size( (0, )) for input in inputs ] return sizes def _profile_single_kernel( self, runner: TunableRunner, inputs: List[torch.Tensor], tactic: Any, tuning_config: TuningConfig, use_cuda_graph: bool = False, **kwargs, ) -> float: """Profile a single kernel implementation for performance measurement. Args: runner (TunableRunner): The runner implementation to profile inputs (List[torch.Tensor]): Input tensors for the kernel tactic (Any): Tactic to use for this profiling run Returns: Average execution time in milliseconds Note: The method performs warmup runs, then measures multiple iterations to get an average execution time. Stream synchronization and delays are used to ensure accurate timing. """ input_tensor_batches = self._prepare_input_tensors_with_batches( inputs, tuning_config) stream = torch.cuda.current_stream() # If the warm up time is longer than 0.5ms, we will profile the kernel with fewer repeats. profile_fewer_repeat = 2 short_profile_threshold_ms = 1 avg_time = float('inf') def pure_profile(stream: torch.cuda.Stream, repeat: int): start = torch.cuda.Event(enable_timing=True) end = torch.cuda.Event(enable_timing=True) graph = torch.cuda.CUDAGraph() with torch.cuda.stream(stream): if use_cuda_graph: with torch.cuda.graph(graph): for r in range(repeat): runner( input_tensor_batches[r % len(input_tensor_batches)], tactic=tactic, **kwargs, ) stream.synchronize() # Delay the profiled kernel launch to eliminate affects of host time overhead in profiling. # TODO: This is build time sensitive, O(tactic_num * impl_num * num_profile * tunable_ops) # Consider apply a preprofiling to estimate the kernel execution time, then decide the necessity. if use_cuda_graph: delay_kernel(self._CUDA_GRAPH_DELAY_MICRO_SECS, stream) else: delay_kernel(self.stream_delay_micro_secs, stream) start.record() if use_cuda_graph: graph.replay() else: for r in range(repeat): runner( input_tensor_batches[r % len(input_tensor_batches)], tactic=tactic, **kwargs, ) end.record() stream.synchronize() return start.elapsed_time(end) / repeat for _ in range(self.warmup): runner(input_tensor_batches[-1], tactic=tactic, **kwargs) fewer_repeat_avg_time = pure_profile(stream, profile_fewer_repeat) disable_short_profile = os.environ.get( "TLLM_AUTOTUNER_DISABLE_SHORT_PROFILE", "0") == "1" if fewer_repeat_avg_time > short_profile_threshold_ms and not disable_short_profile: # directly use the few repeat estimated time to avoid redundant profiling avg_time = fewer_repeat_avg_time else: # profile the kernel with the full repeat to get precise time avg_time = pure_profile(stream, self.repeat) shapes = self._get_input_sizes(inputs) self._debug_logger( f"[Autotuner] Profiled runner={runner}, tactic={tactic}, shapes={shapes}: {avg_time:.6f}ms." ) return avg_time def _optimization_profiles( self, tuning_config: TuningConfig, inputs: List[torch.Tensor]) -> List[OptimizationProfile]: """Generate optimization profiles for autotuning. Args: tuning_config (TuningConfig): Tuning configuration inputs (List[torch.Tensor]): List of input tensors Returns: List of OptimizationProfile objects representing different configurations Note: This method performs a cartesian product of all possible dimension combinations specified in dynamic_tensor_specs. """ # every dimension created from the concrete input tensor shape # generate some dynamic dimension description based on the dynamic_tensors # Zero handles the case where a TRTLLM op has optional or scalar inputs. base_profile = OptimizationProfile( [[StaticDim(x) for x in t.size()] if isinstance(t, torch.Tensor) else [StaticDim(0)] for t in inputs]) generated_profiles: List[OptimizationProfile] = [] dynamic_dims = [] for spec in tuning_config.dynamic_tensor_specs: assert callable(spec.gen_tuning_buckets) or isinstance(spec.gen_tuning_buckets, (list, tuple)), \ "The given dynamic dimension must provide a opt value generation function or a list of opt values" if callable(spec.gen_tuning_buckets): if tuning_config.tune_max_num_tokens is None: # Use the current input size as the opt value opt_shapes = spec.gen_tuning_buckets( base_profile.shapes[spec.input_idx][spec.dim_idx].val) else: # Use the tune_max_num_tokens as the opt value opt_shapes = spec.gen_tuning_buckets( tuning_config.tune_max_num_tokens) else: # Default values is an empty tuple, means that user does not want to tune this dimension. opt_shapes = spec.gen_tuning_buckets # Add the current input value as one of the opt values opt_shapes = set(opt_shapes) opt_shapes.add( spec.map_to_tuning_buckets( base_profile.shapes[spec.input_idx][spec.dim_idx].val)) opt_shapes = sorted(list(opt_shapes)) opt_shapes_max = tuple(opt_shapes[1:]) + (float('inf'), ) opt_shapes_max = { v1: v2 for v1, v2 in zip(opt_shapes, opt_shapes_max) } dynamic_dims.append( (spec.input_idx, spec.dim_idx, opt_shapes_max, opt_shapes)) # grid search, do cartesian product for all the dynamic axis dim_grids = itertools.product(*[d[-1] for d in dynamic_dims]) for opt_point in dim_grids: p = copy.deepcopy(base_profile) for pos, (input_idx, dim_idx, opt_shapes_max, opt_shapes) in enumerate(dynamic_dims): opt_value = opt_point[pos] #TODO: fix me, how to set the min and max? min_value = opt_value max_value = opt_shapes_max[opt_value] p.shapes[input_idx][dim_idx] = DynamicDim( min_value, opt_value, max_value) # Adjust the profile to satisfy the constraints for spec in tuning_config.constraint_specs: min_value = opt_value = max_value = spec.infer_shape( p.get_opt_shapes()) if p.shapes[spec.input_idx] == [StaticDim(0)]: continue p.shapes[spec.input_idx][spec.dim_idx] = DynamicDim( min_value, opt_value, max_value) generated_profiles.append(p) self._debug_logger(f"[Autotuner] Generated profile: {p}") return generated_profiles @classmethod @lru_cache(maxsize=None) def _find_nearest_profile( cls, shapes: Tuple[torch.Size], dynamic_tensor_specs: Tuple[DynamicTensorSpec, ...], constraint_specs: Tuple[ConstraintSpec, ...], tune_max_num_tokens: int = None, ) -> Tuple: """Find the nearest optimization profile for given inputs User can define their own nearest profile generation method to reduce the host overhead. Args: shapes: Tuple of input tensor shapes tuning_config: Tuning configuration Return: Tuple: A tuple containing: - attributes: Tuple of runner attributes, sorted. - profile: Tuple of input tensor shapes """ base_profile = list(list(shape) for shape in shapes) for spec in dynamic_tensor_specs: base_profile[spec.input_idx][ spec.dim_idx] = spec.map_to_tuning_buckets( base_profile[spec.input_idx][spec.dim_idx]) if tune_max_num_tokens is not None: base_profile[spec.input_idx][spec.dim_idx] = min( base_profile[spec.input_idx][spec.dim_idx], tune_max_num_tokens) # associated dimensions dependent on other free dynamic dimensions, so assign -1 in the profile for spec in constraint_specs: if base_profile[spec.input_idx] == [0]: continue base_profile[spec.input_idx][spec.dim_idx] = -1 return tuple(tuple(shape) for shape in base_profile) def _create_tensor_like(self, origin_tensor: torch.Tensor, dims: List[Dim]) -> torch.Tensor: """Create a new tensor matching the properties of the original tensor. Args: origin_tensor (torch.Tensor): Template tensor to match dims (List[Dim]): List of dimensions for the new tensor Returns: New tensor with specified dimensions and matching properties Note: Creates a zero tensor with the same dtype and device as the original, but with dimensions specified by the dims parameter. """ dtype = origin_tensor.dtype device = origin_tensor.device shapes = [] for i, d in enumerate(dims): if isinstance(d, StaticDim): assert d.val == origin_tensor.shape[i] shapes.append(d.val) else: # TODO: how to make sure the created Tensor has the min/max info assert isinstance(d, DynamicDim) shapes.append(d.opt) if len(dims) == 2 and isinstance(dims[0], DynamicDim) and isinstance( dims[1], StaticDim) and (dtype == torch.int32 or dtype == torch.int64): # We should be carefully about int values, since they might be index like topk_index. # We want to keep them legal, so just repeating input tensor. repeat_times = (shapes[0] + origin_tensor.shape[0] - 1) // origin_tensor.shape[0] dup_tensor = origin_tensor.repeat(repeat_times, 1)[:shapes[0]] return dup_tensor # TODO: FIXME, sometimes the content of the tensor can affect the performance, like MOE # One solution is to manituplate the tensor content to make it more like the real data # during the tuning process. This can by controlled in the preparation phase by the runner. # It must not use all zero tensors. Otherwise the timing results become unreliable. if dtype == torch.float4_e2m1fn_x2: return torch.randint(-5, 5, shapes, device=device).to(torch.uint8).view(dtype) else: return torch.randint(-5, 5, shapes, device=device).to(dtype) def _prepare_input_tensors( self, profile: OptimizationProfile, inputs: List[torch.Tensor]) -> List[torch.Tensor]: tensors = [] for i, p in enumerate(profile.shapes): if any(isinstance(d, DynamicDim) for d in p): tensor = self._create_tensor_like(inputs[i], p) else: tensor = inputs[i] tensors.append(tensor) return tensors def _prepare_input_tensors_with_batches( self, inputs: List[torch.Tensor], tuning_config: TuningConfig, ) -> List[List[torch.Tensor]]: if not tuning_config.use_cold_l2_cache: return [inputs] one_buffer_bytes = sum( input.numel() * input.element_size() if isinstance(input, torch.Tensor) else 0 for input in inputs) if one_buffer_bytes <= 0: self._debug_logger( "[Autotuner] No tensor inputs or zero-sized tensors; falling back to single-batch profiling." ) return [inputs] num_buffers = self._get_l2_cache_size_in_bytes( ) * 3 // one_buffer_bytes + 1 num_buffers = min(num_buffers, self.repeat + 1) inputs_list = [inputs] for _ in range(num_buffers - 1): inputs_list.append( list(t.clone() if isinstance(t, torch.Tensor) else t for t in inputs)) self._debug_logger( f"[Autotuner] use_cold_l2_cache={tuning_config.use_cold_l2_cache}, use {num_buffers} different tensors for profiling" ) return inputs_list def clear_cache(self) -> None: """Clear the profiling cache.""" self.profiling_cache.clear() def reset_statistics(self) -> None: """Reset all statistics counters.""" self.stats = AutoTunerStatistics() def print_profiling_cache(self): self._debug_logger(f"[Autotuner] The profiling_cache entries:") self._debug_logger( f"[Autotuner] Cache contents: (custom_op, runner, hash(attributes), shape_profiles) -> (runner_id, tactic, shape_profile(ignored))" ) for key, value in self.profiling_cache.cache.items(): runner_id, tactic, min_time = value self._debug_logger( f"[Autotuner] {key}: (runner_id={runner_id}, tactic={tactic}, min_time={min_time})" ) self.print_statistics() def print_statistics(self): self._debug_logger(f"[Autotuner] The statistics:") for line in self.stats.__str__().split("\n"): self._debug_logger(line) @contextlib.contextmanager def capture(self): """Context manager for capturing execution contexts for testing. Returns a TacticsCapture object that can be iterated to get all valid (runner, tactic) combinations. Example: >>> # Single context case >>> with AutoTuner.get().capture() as tactics_capture: ... y = custom_op.forward(x) >>> >>> for runner, tactic in tactics_capture: ... with AutoTuner.get().replay(runner, tactic): ... y = custom_op.forward(x) >>> # Multiple contexts case >>> with AutoTuner.get().capture() as tactics_capture: ... y = custom_op1.forward(x) ... z = custom_op2.forward(y) >>> >>> for config in tactics_capture: ... with AutoTuner.get().replay(config): ... y = custom_op1.forward(x) ... z = custom_op2.forward(y) """ tactics_capture = self.TacticsCapture(self) self._active_capture = tactics_capture try: yield tactics_capture finally: self._active_capture = None self._last_capture = tactics_capture @contextlib.contextmanager def replay(self, *config: Tuple[Tuple[TunableRunner, int], ...]): """Context manager for replaying with specific runner/tactic configuration. Args: config: - A tuple of (runner, tactic) pairs. The tuple size matches the number of captured choose_one() contexts. """ # Parse config argument if len(config) == 1: if isinstance(config[0], tuple): # Multiple contexts: replay(((r0,t0), (r1,t1), ...)) runner_tactic_pairs = list(config[0]) else: # Also handle single context passed as replay((runner, tactic)) runner_tactic_pairs = [config[0]] else: raise ValueError( f"Invalid config for replay: {config}\n" "Expected replay(((runner, tactic), (runner, tactic), ...))") # Find the TacticsCapture to use tactics_capture = self._active_capture or self._last_capture if tactics_capture is None: raise RuntimeError( "No TacticsCapture available for replay. " "Make sure you've called capture() before replay().") # Temporarily set as active capture during replay prev_active = self._active_capture self._active_capture = tactics_capture runner_tactic_list = [] for ctx_idx, (runner, tactic) in enumerate(runner_tactic_pairs): runners = tactics_capture._captured_contexts[ctx_idx]['runners'] runner_idx = runners.index(runner) runner_tactic_list.append((runner_idx, tactic)) self._debug_logger( f"[Autotuner][replay]: Testing configuration: {runner_tactic_list}") # Replay the contexts with given (runner, tactic) pairs tactics_capture._replay_runner_tactic_list = runner_tactic_list tactics_capture._replay_context_idx = 0 try: yield finally: tactics_capture._replay_runner_tactic_list = None tactics_capture._replay_context_idx = 0 # Restore previous active capture state self._active_capture = prev_active def _get_l2_cache_size_in_bytes(self, device_id: int = 0) -> int: device = self._checkCudaErrors(driver.cuDeviceGet(device_id)) return self._checkCudaErrors( driver.cuDeviceGetAttribute( driver.CUdevice_attribute.CU_DEVICE_ATTRIBUTE_L2_CACHE_SIZE, device, )) def _checkCudaErrors(self, result) -> Any: status = result[0] if status != driver.CUresult.CUDA_SUCCESS: code = getattr(status, "value", status) raise RuntimeError( f"CUDA error code={code}({self._cudaGetErrorEnum(status)})") # CUDA APIs always return the status as the first element of the result tuple if len(result) == 1: return None elif len(result) == 2: return result[1] else: return result[1:] def _cudaGetErrorEnum(self, error) -> str: from cuda.bindings import nvrtc if isinstance(error, driver.CUresult): err, name = driver.cuGetErrorName(error) return name if err == driver.CUresult.CUDA_SUCCESS else "" elif isinstance(error, nvrtc.nvrtcResult): return nvrtc.nvrtcGetErrorString(error)[1] else: raise RuntimeError("Unknown error type: {}".format(error))