# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project import itertools from collections.abc import Iterable, Mapping from dataclasses import dataclass, field from typing import Any, Literal, Protocol, overload import torch import torch.nn as nn from torch.func import functional_call from transformers import PretrainedConfig from typing_extensions import deprecated import vllm.envs as envs from vllm.config import VllmConfig from vllm.distributed import ( get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size, ) from vllm.logger import init_logger from vllm.model_executor.model_loader.weight_utils import default_weight_loader from vllm.multimodal import NestedTensors from vllm.sequence import IntermediateTensors from vllm.utils import ( cdiv, is_pin_memory_available, is_uva_available, ) from vllm.utils.torch_utils import ( direct_register_custom_op, get_cuda_view_from_cpu_tensor, ) logger = init_logger(__name__) WeightsMapping = Mapping[str, str | None] """If a key maps to a value of `None`, the corresponding weight is ignored.""" @dataclass class WeightsMapper: """Maps the name of each weight if they match the following patterns.""" orig_to_new_substr: WeightsMapping = field(default_factory=dict) orig_to_new_prefix: WeightsMapping = field(default_factory=dict) orig_to_new_suffix: WeightsMapping = field(default_factory=dict) def _map_name(self, key: str) -> str | None: for substr, new_key in self.orig_to_new_substr.items(): if substr in key: if new_key is None: return None key = key.replace(substr, new_key, 1) for prefix, new_key in self.orig_to_new_prefix.items(): if key.startswith(prefix): if new_key is None: return None key = key.replace(prefix, new_key, 1) for suffix, new_key in self.orig_to_new_suffix.items(): if key.endswith(suffix): if new_key is None: return None key = new_key.join(key.rsplit(suffix, 1)) return key def apply( self, weights: Iterable[tuple[str, torch.Tensor]] ) -> Iterable[tuple[str, torch.Tensor]]: return ( (out_name, data) for name, data in weights if (out_name := self._map_name(name)) is not None ) def apply_list(self, values: list[str]) -> list[str]: return [ out_name for name in values if (out_name := self._map_name(name)) is not None ] def apply_dict(self, values: dict[str, Any]) -> dict[str, Any]: return { out_name: value for name, value in values.items() if (out_name := self._map_name(name)) is not None } class AutoWeightsLoader: """ Helper class to load weights into a [`torch.nn.Module`][]. It is able to automatically detect child modules and parameters while iterating over the weights only once. The weight loading logic for individual modules can be overridden by defining a `load_weights` method. Similarly, the weight loading logic for individual parameters can be overridden by defining a `weight_loader` method. Detailed weight loading information can be viewed by setting the environment variable `VLLM_LOGGING_LEVEL=DEBUG`. """ # Models trained using early version ColossalAI # may include these tensors in checkpoint. Skip them. ROTARY_EMBEDS_UNUSED_WEIGHTS = [ "rotary_emb.inv_freq", "rotary_emb.cos_cached", "rotary_emb.sin_cached", ] def __init__( self, module: nn.Module, *, skip_prefixes: list[str] | None = None, skip_substrs: list[str] | None = None, ignore_unexpected_prefixes: list[str] | None = None, ignore_unexpected_suffixes: list[str] | None = None, ) -> None: super().__init__() self.module = module self.skip_prefixes = skip_prefixes or [] self.skip_substrs = skip_substrs or [] self.ignore_unexpected_prefixes = ignore_unexpected_prefixes or [] self.ignore_unexpected_suffixes = ignore_unexpected_suffixes or [] # update default skip_substrs self.skip_substrs += self.ROTARY_EMBEDS_UNUSED_WEIGHTS def _groupby_prefix( self, weights: Iterable[tuple[str, torch.Tensor]], ) -> Iterable[tuple[str, Iterable[tuple[str, torch.Tensor]]]]: weights_by_parts = ( (weight_name.split(".", 1), weight_data) for weight_name, weight_data in weights ) for prefix, group in itertools.groupby(weights_by_parts, key=lambda x: x[0][0]): yield ( prefix, # Because maxsplit=1 in weight_name.split(...), # the length of `parts` must either be 1 or 2 ( ("" if len(parts) == 1 else parts[1], weights_data) for parts, weights_data in group ), ) def _get_qualname(self, prefix: str, rest: str) -> str: if prefix == "": return rest if rest == "": return prefix return ".".join((prefix, rest)) def _can_skip(self, qualname: str) -> bool: return any(qualname.startswith(p) for p in self.skip_prefixes) or any( substr in qualname for substr in self.skip_substrs ) def _can_ignore_unexpected(self, qualname: str) -> bool: iup = (qualname.startswith(p) for p in self.ignore_unexpected_prefixes) ius = (qualname.endswith(s) for s in self.ignore_unexpected_suffixes) return any(iup) or any(ius) def _load_param( self, base_prefix: str, param: nn.Parameter, weights: Iterable[tuple[str, torch.Tensor]], ) -> Iterable[str]: for weight_name, weight_data in weights: weight_qualname = self._get_qualname(base_prefix, weight_name) if self._can_skip(weight_qualname): logger.debug("Skipping weight %s", weight_qualname) continue if weight_name != "": if self._can_ignore_unexpected(weight_qualname): logger.debug("Ignoring weight %s", weight_qualname) continue raise ValueError( f"Attempted to load nested weight '{weight_qualname}' " f"into a single parameter '{base_prefix}'" ) weight_loader = getattr(param, "weight_loader", default_weight_loader) weight_loader(param, weight_data) logger.debug("Loaded weight %s with shape %s", weight_qualname, param.shape) yield weight_qualname def _add_loadable_non_param_tensors( self, module: nn.Module, child_params: dict[str, torch.Tensor] ): """ Add tensor names that are not in the model params that may be in the safetensors, e.g., batch normalization stats. """ if isinstance( module, ( nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d, nn.LazyBatchNorm1d, nn.LazyBatchNorm2d, nn.LazyBatchNorm3d, nn.SyncBatchNorm, ), ): module_state_dict = module.state_dict() for stat_name in ("running_mean", "running_var", "num_batches_tracked"): child_params[stat_name] = module_state_dict[stat_name] def _load_module( self, base_prefix: str, module: nn.Module, weights: Iterable[tuple[str, torch.Tensor]], ) -> Iterable[str]: if isinstance(module, PPMissingLayer): return # Avoid infinite recursion since this function is typically # called inside load_weights of the module itself if module != self.module: module_load_weights = getattr(module, "load_weights", None) if callable(module_load_weights): loaded_params = module_load_weights(weights) if loaded_params is None: logger.warning( "Unable to collect loaded parameters for module %s", module ) else: yield from map( lambda x: self._get_qualname(base_prefix, x), loaded_params, ) child_modules = dict(module.named_children()) child_params = dict(module.named_parameters(recurse=False)) # Add missing tensors the weight loader needs to be able to load # that aren't registered as params, e.g., batchnorm statistics. self._add_loadable_non_param_tensors(module, child_params) for child_prefix, child_weights in self._groupby_prefix(weights): prefix = self._get_qualname(base_prefix, child_prefix) if child_prefix in child_modules: if self._can_skip(prefix + "."): logger.debug("Skipping module %s", prefix) continue yield from self._load_module( prefix, child_modules[child_prefix], child_weights ) elif child_prefix in child_params: if self._can_skip(prefix): logger.debug("Skipping param %s", prefix) continue yield from self._load_param( prefix, child_params[child_prefix], child_weights ) else: can_skip_module = self._can_skip(prefix + ".") can_skip_param = self._can_skip(prefix) if can_skip_module or can_skip_param: logger.debug("Skipping missing %s", prefix) continue can_ignore_module = self._can_ignore_unexpected(prefix + ".") can_ignore_param = self._can_ignore_unexpected(prefix) if can_ignore_module or can_ignore_param: logger.debug("Ignoring missing %s", prefix) continue msg = ( f"There is no module or parameter named '{prefix}' " f"in {type(self.module).__name__}" ) raise ValueError(msg) def load_weights( self, weights: Iterable[tuple[str, torch.Tensor]], *, mapper: WeightsMapper | None = None, ) -> set[str]: if mapper is not None: weights = mapper.apply(weights) # filter out weights with first-prefix/substr to skip in name weights = ( (name, weight) for name, weight in weights if not self._can_skip(name) ) autoloaded_weights = set(self._load_module("", self.module, weights)) return autoloaded_weights def init_vllm_registered_model( vllm_config: VllmConfig, *, prefix: str = "", hf_config: PretrainedConfig | None = None, architectures: list[str] | None = None, ) -> nn.Module: """ Helper function to initialize an inner model registered to vLLM, based on the arguments passed to the outer vLLM model. """ from vllm.model_executor.model_loader.utils import initialize_model if hf_config is None and architectures is not None: # So that the architectures field is overridden hf_config = vllm_config.model_config.hf_config if hf_config is not None: vllm_config = vllm_config.with_hf_config(hf_config, architectures=architectures) return initialize_model(vllm_config=vllm_config, prefix=prefix) @overload def flatten_bn(x: torch.Tensor) -> torch.Tensor: ... @overload def flatten_bn(x: list[torch.Tensor]) -> list[torch.Tensor]: ... @overload def flatten_bn( x: list[torch.Tensor] | torch.Tensor, *, concat: Literal[True], ) -> torch.Tensor: ... @overload def flatten_bn( x: list[torch.Tensor] | torch.Tensor, *, concat: bool = False, ) -> list[torch.Tensor] | torch.Tensor: ... def flatten_bn( x: list[torch.Tensor] | torch.Tensor, *, concat: bool = False, ) -> list[torch.Tensor] | torch.Tensor: """ Flatten the `B` and `N` dimensions of batched multimodal inputs. The input tensor should have shape `(B, N, ...)`. """ if isinstance(x, torch.Tensor): return x.flatten(0, 1) if concat: return torch.cat(x) return [x_n for x_b in x for x_n in x_b] def _flatten_embeddings(embeddings: NestedTensors) -> torch.Tensor: """ Recursively flattens and concatenates NestedTensors on all but the last dimension. """ if isinstance(embeddings, torch.Tensor): # Flatten all but the last dimension. return embeddings.flatten(0, -2) return torch.cat(tuple(_flatten_embeddings(t) for t in embeddings)) def _embedding_count_expression(embeddings: NestedTensors) -> str: """ Constructs a debugging representation of the number of embeddings in the NestedTensors. """ if isinstance(embeddings, torch.Tensor): return " x ".join([str(dim) for dim in embeddings.shape[:-1]]) return " + ".join(_embedding_count_expression(inner) for inner in embeddings) def split_list_into_ranges(lst: torch.Tensor, interval: int) -> list[list[int]]: ranges: list[list[int]] = [[] for _ in range((max(lst) // interval) + 1)] for num in lst: index = num // interval ranges[index].append(num) return ranges def _merge_multimodal_embeddings( inputs_embeds: torch.Tensor, multimodal_embeddings: NestedTensors, is_multimodal: torch.Tensor, ) -> torch.Tensor: """ Merge `multimodal_embeddings` into `inputs_embeds` by overwriting the positions in `inputs_embeds` corresponding to placeholder tokens in `input_ids`. Note: This updates `inputs_embeds` in place. """ if len(multimodal_embeddings) == 0: return inputs_embeds mm_embeds_flat = _flatten_embeddings(multimodal_embeddings) input_dtype = inputs_embeds.dtype try: # For debugging # inputs_embeds[is_multimodal] = mm_embeds_flat.to(dtype=input_dtype) # NOTE: This can avoid D2H sync (#22105), but fails to # raise an error if is_multimodal.sum() < len(mm_embeds_flat) inputs_embeds.masked_scatter_( is_multimodal.unsqueeze(-1), mm_embeds_flat.to(dtype=input_dtype) ) except RuntimeError as e: num_actual_tokens = len(mm_embeds_flat) num_expected_tokens = is_multimodal.sum().item() if num_actual_tokens != num_expected_tokens: expr = _embedding_count_expression(multimodal_embeddings) raise ValueError( f"Attempted to assign {expr} = {num_actual_tokens} " f"multimodal tokens to {num_expected_tokens} placeholders" ) from e raise ValueError("Error during masked scatter operation") from e return inputs_embeds @deprecated( "`merge_multimodal_embeddings` has been replaced with " "`SupportsMultiModal.get_input_embeddings` and will be " "removed in v0.12." ) def merge_multimodal_embeddings( input_ids: torch.Tensor, inputs_embeds: torch.Tensor, multimodal_embeddings: NestedTensors, placeholder_token_id: int | list[int], ) -> torch.Tensor: """ Merge `multimodal_embeddings` into `inputs_embeds` by overwriting the positions in `inputs_embeds` corresponding to placeholder tokens in `input_ids`. `placeholder_token_id` can be a list of token ids (e.g, token ids of img_start, img_break, and img_end tokens) when needed: This means the order of these tokens in the `input_ids` MUST MATCH the order of their embeddings in `multimodal_embeddings` since we need to slice-merge instead of individually scattering. For example, if input_ids is "TTTTTSIIIBIIIBIIIETTT", where - T is text token - S is image start token - I is image embedding token - B is image break token - E is image end token. Then the image embeddings (that correspond to I's) from vision encoder must be padded with embeddings of S, B, and E in the same order of input_ids for a correct embedding merge. Note: This updates `inputs_embeds` in place. """ if isinstance(placeholder_token_id, list): is_multimodal = isin_list(input_ids, placeholder_token_id) else: is_multimodal = input_ids == placeholder_token_id return _merge_multimodal_embeddings( inputs_embeds, multimodal_embeddings=multimodal_embeddings, is_multimodal=is_multimodal, ) def isin_list( elements: torch.Tensor, test_elements_list: list[int], ) -> torch.Tensor: test_elements = torch.tensor( test_elements_list, pin_memory=is_pin_memory_available(), ).to(device=elements.device, non_blocking=True) return torch.isin(elements, test_elements) class LayerFn(Protocol): def __call__(self, prefix: str) -> torch.nn.Module: ... class PPMissingLayer(torch.nn.Identity): """ A placeholder layer for missing layers in a pipeline parallel model. """ def __init__(self, *args, **kwargs): super().__init__() def forward(self, *args, **kwargs): """Return the first arg from args or the first value from kwargs.""" return args[0] if args else next(iter(kwargs.values())) _CPU_OFFLOAD_BYTES = 0 _CPU_OFFLOAD_MAX_BYTES = 0 def set_cpu_offload_max_bytes(max_bytes: int) -> None: global _CPU_OFFLOAD_MAX_BYTES, _CPU_OFFLOAD_BYTES _CPU_OFFLOAD_BYTES = 0 _CPU_OFFLOAD_MAX_BYTES = max_bytes def maybe_offload_to_cpu(module: torch.nn.Module) -> torch.nn.Module: if (params := next(module.parameters(), None)) is None: return module device = params.device if device == torch.device("cpu"): return module global _CPU_OFFLOAD_MAX_BYTES, _CPU_OFFLOAD_BYTES if _CPU_OFFLOAD_BYTES >= _CPU_OFFLOAD_MAX_BYTES: return module pin_memory = is_pin_memory_available() uva_available = is_uva_available() if envs.VLLM_USE_V1: assert uva_available, "V1 CPU offloading requires uva (pin memory) support" uva_offloading = True else: uva_offloading = False # offload parameters to CPU # use pin_memory if possible, which helps cudagraph capture speed offloaded_parameters = False for p in module.parameters(): if _CPU_OFFLOAD_BYTES >= _CPU_OFFLOAD_MAX_BYTES: # we use per-parameter offloading # one module might have some parameters offloaded and some not break # `torch.empty_like` does not support `pin_memory` argument cpu_data = torch.empty_strided( size=p.data.size(), stride=p.data.stride(), dtype=p.data.dtype, layout=p.data.layout, device="cpu", pin_memory=pin_memory, ) cpu_data.copy_(p.data) if not uva_offloading: p.data = cpu_data else: # keep the cpu data alive p._vllm_offloaded_cpu_data = cpu_data p.data = get_cuda_view_from_cpu_tensor(cpu_data) _CPU_OFFLOAD_BYTES += p.data.numel() * p.data.element_size() offloaded_parameters = True if offloaded_parameters and not uva_offloading: original_forward = module.forward def forward(*args, **kwargs): module.forward = original_forward device_state = { # here we blindly call `to(device)` # if the parameter is already on the device, it will be a no-op k: v.to(device, non_blocking=True) for k, v in module.state_dict().items() } output = functional_call(module, device_state, args=args, kwargs=kwargs) module.forward = forward return output module.forward = forward return module def make_layers( num_hidden_layers: int, layer_fn: LayerFn, prefix: str, ) -> tuple[int, int, torch.nn.ModuleList]: """Make a list of layers with the given layer function, taking pipeline parallelism into account. """ from vllm.distributed.parallel_state import get_pp_group from vllm.distributed.utils import get_pp_indices start_layer, end_layer = get_pp_indices( num_hidden_layers, get_pp_group().rank_in_group, get_pp_group().world_size ) modules = torch.nn.ModuleList( [PPMissingLayer() for _ in range(start_layer)] + [ maybe_offload_to_cpu(layer_fn(prefix=f"{prefix}.{idx}")) for idx in range(start_layer, end_layer) ] + [PPMissingLayer() for _ in range(end_layer, num_hidden_layers)] ) return start_layer, end_layer, modules # NOTE: don't use lru_cache here because it can prevent garbage collection _model_to_pp_missing_layer_names: dict[int, list[str]] = {} def get_pp_missing_layer_names(model: torch.nn.Module) -> list[str]: """Get the names of the missing layers in a pipeline parallel model.""" model_id = id(model) if model_id in _model_to_pp_missing_layer_names: return _model_to_pp_missing_layer_names[model_id] missing_layer_names = [] for name, module in model.named_modules(): if isinstance(module, PPMissingLayer): # NOTE: the trailing dot is used to match the prefix of the layer. # without the dot, we could match a layer that is not missing, # e.g., 'encoder.layer.1' would match 'encoder.layer.11' missing_layer_names.append(name + ".") _model_to_pp_missing_layer_names[model_id] = missing_layer_names return missing_layer_names def is_pp_missing_parameter(name: str, model: torch.nn.Module) -> bool: """Check if a parameter is missing in a pipeline parallel model.""" if isinstance(model, PPMissingLayer): return True return any( name.startswith(missing_layer_name) for missing_layer_name in get_pp_missing_layer_names(model) ) def make_empty_intermediate_tensors_factory(keys: list[str], hidden_size: int): def make_empty_intermediate_tensors( batch_size: int, dtype: torch.dtype, device: torch.device, ) -> IntermediateTensors: return IntermediateTensors( { key: torch.zeros((batch_size, hidden_size), dtype=dtype, device=device) for key in keys } ) return make_empty_intermediate_tensors def maybe_prefix(prefix: str, name: str) -> str: """Add a prefix to a name if the prefix is non-empty. Args: prefix: The prefix to add. If empty, no prefix will be added. name: The name to potentially prefix. Returns: The string "prefix.name" if prefix was non-empty, otherwise just "name". """ return name if not prefix else f"{prefix}.{name}" def extract_layer_index(layer_name: str, num_attn_module: int = 1) -> int: """ Extract the layer index from the module name. Examples: - "encoder.layers.0" -> 0 - "encoder.layers.1.self_attn" -> 1 - "2.self_attn" -> 2 - "model.encoder.layers.0.sub.1" -> ValueError if num_attn_module == 1 """ subnames = layer_name.split(".") int_vals: list[int] = [] for subname in subnames: try: int_vals.append(int(subname)) except ValueError: continue if num_attn_module == 1 or "attn" not in layer_name: assert len(int_vals) == 1, ( f"layer name {layer_name} should only contain one integer" ) return int_vals[0] else: assert len(int_vals) <= 2, ( f"layer name {layer_name} should contain most two integers" ) layer_index = ( int_vals[0] * num_attn_module + int_vals[1] if len(int_vals) == 2 else int_vals[0] ) return layer_index def cast_overflow_tensors( tensors: torch.Tensor, offset: float = 1000, ) -> torch.Tensor: if tensors.isinf().any() or tensors.isnan().any(): clamp_value = torch.finfo(tensors.dtype).max - offset tensors = torch.clamp(tensors, min=-clamp_value, max=clamp_value) return tensors def fast_topk( values: torch.Tensor, topk: int, dim: int ) -> tuple[torch.Tensor, torch.Tensor]: """ Optimized topk implementation that uses torch.max for k=1 case. This function provides better performance for the common case of k=1 by using torch.max instead of the more general torch.topk. Args: values: Input tensor to find top-k values from topk: Number of top values to return (k). Must be > 0. dim: Dimension along which to compute topk Returns: Tuple of (values, indices) where values are the top-k values and indices are their corresponding indices in the input tensor """ if topk == 1: # Use max along the specified dimension to get both value and index return torch.max(values, dim=dim, keepdim=True) else: # Use topk for efficiency with larger k values return torch.topk(values, topk, dim=dim) # Chunk x along the num_tokens axis for sequence parallelism # NOTE: This is wrapped in a torch custom op to work around the following issue: # The output tensor can have a sequence length 0 at small input sequence lengths # even though we explicitly pad to avoid this. def sequence_parallel_chunk(x: torch.Tensor) -> torch.Tensor: return torch.ops.vllm.sequence_parallel_chunk_impl(x) def sequence_parallel_chunk_impl(x: torch.Tensor) -> torch.Tensor: tp_size = get_tensor_model_parallel_world_size() tp_rank = get_tensor_model_parallel_rank() # all_gather needs the sequence length to be divisible by tp_size seq_len = x.size(0) remainder = seq_len % tp_size if remainder != 0: pad_len = tp_size - remainder y = nn.functional.pad(x, (0, 0, 0, pad_len)) else: y = x chunk = y.shape[0] // tp_size start = tp_rank * chunk return torch.narrow(y, 0, start, chunk) def sequence_parallel_chunk_impl_fake(x: torch.Tensor) -> torch.Tensor: tp_size = get_tensor_model_parallel_world_size() seq_len = cdiv(x.size(0), tp_size) shape = list(x.shape) shape[0] = seq_len out = torch.empty(shape, dtype=x.dtype, device=x.device) return out direct_register_custom_op( op_name="sequence_parallel_chunk_impl", op_func=sequence_parallel_chunk_impl, fake_impl=sequence_parallel_chunk_impl_fake, tags=(torch.Tag.needs_fixed_stride_order,), )