Quelle  transformer.py   Sprache: Python

import json
import logging
import math
from dataclasses import dataclass
from pathlib import Path
from typing import Any, List, Mapping, Optional, Union

import safetensors.torch
import torch
from torch import nn

from mistral_inference.args import PATCH_MERGE, TransformerArgs
from mistral_inference.cache import BufferCache, CacheInputMetadata
from mistral_inference.lora import LoRALoaderMixin
from mistral_inference.model import ModelBase
from mistral_inference.rope import precompute_freqs_cis
from mistral_inference.transformer_layers import RMSNorm, TransformerBlock
from mistral_inference.vision_encoder import PatchMerger, VisionLanguageAdapter, VisionTransformer


@dataclass
class SimpleInputMetadata:
    # rope absolute positions
    positions: torch.Tensor

    @staticmethod
    def from_seqlens(seqlens: List[int], device: torch.device) -> "SimpleInputMetadata":
        return SimpleInputMetadata(
            positions=torch.cat([torch.arange(0, seqlen) for seqlen in seqlens]).to(device=device, dtype=torch.long)
        )


class Transformer(ModelBase, LoRALoaderMixin):
    def __init__(
        self,
        args: TransformerArgs,
        pipeline_rank: int = 0,
        num_pipeline_ranks: int = 1,
        softmax_fp32: bool = True,
    ):
        super().__init__()
        self.args = args
        self.vocab_size = args.vocab_size
        self.n_layers = args.n_layers
        self._precomputed_freqs_cis: Optional[torch.Tensor] = None
        assert self.vocab_size > 0
        assert pipeline_rank < num_pipeline_ranks, (pipeline_rank, num_pipeline_ranks)
        self.pipeline_rank = pipeline_rank
        self.num_pipeline_ranks = num_pipeline_ranks
        self.softmax_fp32 = softmax_fp32

        # Modules specific to some ranks:
        self.tok_embeddings: Optional[nn.Embedding] = None
        self.norm: Optional[RMSNorm] = None
        self.output: Optional[nn.Linear] = None
        if pipeline_rank == 0:
            self.tok_embeddings = nn.Embedding(args.vocab_size, args.dim)

            self.vision_encoder: Optional[VisionTransformer] = None
            self.vision_language_adapter: Optional[VisionLanguageAdapter] = None

            if args.vision_encoder is not None:
                self.vision_encoder = VisionTransformer(args.vision_encoder)
                self.vision_language_adapter = VisionLanguageAdapter(
                    args.vision_encoder.hidden_size, args.dim, args.vision_encoder.adapter_bias
                )

                if args.vision_encoder.add_pre_mm_projector_layer_norm:
                    self.pre_mm_projector_norm = RMSNorm(args.vision_encoder.hidden_size, eps=1e-5)

                if args.vision_encoder.mm_projector_id == PATCH_MERGE:
                    self.patch_merger = PatchMerger(
                        vision_encoder_dim=args.vision_encoder.hidden_size,
                        spatial_merge_size=args.vision_encoder.spatial_merge_size,
                    )

        if pipeline_rank == num_pipeline_ranks - 1:
            self.norm = RMSNorm(args.dim, eps=args.norm_eps)
            self.output = nn.Linear(args.dim, args.vocab_size, bias=False)
        # Initialize all layers but slice off those not of this rank.
        layers = [
            TransformerBlock(
                dim=args.dim,
                hidden_dim=args.hidden_dim,
                n_heads=args.n_heads,
                n_kv_heads=args.n_kv_heads,
                head_dim=args.head_dim,
                norm_eps=args.norm_eps,
                lora=args.lora,
                moe=args.moe,
            )
            for _ in range(args.n_layers)
        ]
        num_layers_per_rank = math.ceil(self.n_layers / self.num_pipeline_ranks)
        offset = self.pipeline_rank * num_layers_per_rank
        end = min(self.n_layers, offset + num_layers_per_rank)
        self.layers = nn.ModuleDict({str(i): layers[i] for i in range(offset, end)})
        self.n_local_layers = len(self.layers)

    @property
    def dtype(self) -> torch.dtype:
        return next(self.parameters()).dtype

    @property
    def device(self) -> torch.device:
        return next(self.parameters()).device

    @property
    def freqs_cis(self) -> torch.Tensor:
        # We cache freqs_cis but need to take care that it is on the right device
        # and has the right dtype (complex64). The fact that the dtype is different
        # from the module's  dtype means we cannot register it as a buffer
        if self._precomputed_freqs_cis is None:
            # default to 10**6
            theta = self.args.rope_theta or 1000000.0
            self._precomputed_freqs_cis = precompute_freqs_cis(self.args.head_dim, 128_000, theta)

        if self._precomputed_freqs_cis.device != self.device:
            self._precomputed_freqs_cis = self._precomputed_freqs_cis.to(device=self.device)
        return self._precomputed_freqs_cis

    def embed_vision_language_features(self, input_ids: torch.Tensor, images: List[torch.Tensor]) -> torch.Tensor:
        assert self.tok_embeddings is not None
        assert self.vision_encoder is not None
        assert self.vision_language_adapter is not None
        assert self.args.vision_encoder is not None

        text_locations = input_ids != self.args.vision_encoder.image_token_id
        image_locations = input_ids == self.args.vision_encoder.image_token_id
        text_features = self.tok_embeddings(input_ids[text_locations])

        image_features = self.vision_encoder(images)

        if self.args.vision_encoder.add_pre_mm_projector_layer_norm:
            image_features = self.pre_mm_projector_norm(image_features)

        if self.args.vision_encoder.mm_projector_id == PATCH_MERGE:
            patch_size = self.args.vision_encoder.patch_size
            img_patch_dims = [(img.shape[1] // patch_size, img.shape[2] // patch_size) for img in images]
            image_features = self.patch_merger(image_features, image_sizes=img_patch_dims)

        image_features = self.vision_language_adapter(image_features)

        N_txt, D_txt = text_features.shape
        N_img, D_img = image_features.shape

        seq_len = input_ids.shape[0]

        assert D_txt == D_img, f"Text features dim {D_txt} should be equal to image features dim {D_img}"
        assert seq_len == N_txt + N_img, (
            f"seq_len {seq_len} should be equal to N_txt + N_img {(N_txt, N_img, image_locations.sum().item())}"
        )

        combined_features = torch.empty(
            (seq_len, D_txt),
            dtype=text_features.dtype,
            device=text_features.device,
        )
        combined_features[text_locations, :] = text_features
        combined_features[image_locations, :] = image_features
        return combined_features

    def forward_partial(
        self,
        input_ids: torch.Tensor,
        seqlens: List[int],
        cache: Optional[BufferCache] = None,
        images: Optional[List[torch.Tensor]] = None,
    ) -> torch.Tensor:
        """Local forward pass.

        If doing pipeline parallelism, this will return the activations of the last layer of this stage.
        For the last stage, this will return the normalized final embeddings.
        """
        assert len(seqlens) <= self.args.max_batch_size, (
            f"Max batch size is {self.args.max_batch_size}, got batch size of {len(seqlens)}"
        )
        (num_toks,) = input_ids.shape
        assert sum(seqlens) == num_toks, (sum(seqlens), num_toks)

        input_metadata: List[CacheInputMetadata] | List[SimpleInputMetadata]

        if cache is not None:
            input_metadata = cache.get_input_metadata(seqlens)
        else:
            input_metadata = [SimpleInputMetadata.from_seqlens(seqlens, self.device) for _ in range(len(self.layers))]

        if self.pipeline_rank == 0:
            assert self.tok_embeddings is not None
            if self.vision_encoder is not None and images:
                h = self.embed_vision_language_features(input_ids, images)
            else:
                h = self.tok_embeddings(input_ids)
        else:
            h = torch.empty(num_toks, self.args.dim, device=self.device, dtype=self.dtype)
            torch.distributed.recv(h, src=self.pipeline_rank - 1)

        # freqs_cis is always the same for every layer
        freqs_cis = self.freqs_cis[input_metadata[0].positions]

        for local_layer_id, layer in enumerate(self.layers.values()):
            if cache is not None:
                assert input_metadata is not None
                cache_metadata = input_metadata[local_layer_id]
                assert isinstance(cache_metadata, CacheInputMetadata)
                cache_view = cache.get_view(local_layer_id, cache_metadata)
            else:
                cache_view = None
            h = layer(h, freqs_cis, cache_view)

        if cache is not None:
            cache.update_seqlens(seqlens)
        if self.pipeline_rank < self.num_pipeline_ranks - 1:
            torch.distributed.send(h, dst=self.pipeline_rank + 1)
            return h
        else:
            # Last rank has a final normalization step.
            assert self.norm is not None
            return self.norm(h)  # type: ignore

    def forward(
        self,
        input_ids: torch.Tensor,
        seqlens: List[int],
        cache: Optional[BufferCache] = None,
        images: Optional[List[torch.Tensor]] = None,
    ) -> torch.Tensor:
        h = self.forward_partial(input_ids, seqlens, cache=cache, images=images)
        if self.pipeline_rank < self.num_pipeline_ranks - 1:
            # ignore the intermediate activations as we'll get the final output from
            # the last stage
            outs = torch.empty(h.shape[0], self.vocab_size, device=h.device, dtype=h.dtype)
        else:
            assert self.output is not None
            outs = self.output(h)
        if self.num_pipeline_ranks > 1:
            torch.distributed.broadcast(outs, src=self.num_pipeline_ranks - 1)

        if self.softmax_fp32:
            return outs.float()
        else:
            return outs

    def load_state_dict(self, state_dict: Mapping[str, Any], strict: bool = True, assign: bool = False) -> None:
        state_to_load = {}
        skipped = set([])
        for k, v in state_dict.items():
            if k.startswith("tok_embeddings"):
                if self.pipeline_rank == 0:
                    state_to_load[k] = v
                else:
                    logging.debug(
                        "Skipping parameter %s at pipeline rank %d",
                        k,
                        self.pipeline_rank,
                    )
                    skipped.add(k)
            elif k.startswith("norm") or k.startswith("output"):
                if self.pipeline_rank == self.num_pipeline_ranks - 1:
                    state_to_load[k] = v
                else:
                    logging.debug(
                        "Skipping parameter %s at pipeline rank %d",
                        k,
                        self.pipeline_rank,
                    )
                    skipped.add(k)
            elif k.startswith("layers"):
                layer_id = k.split(".")[1]
                if layer_id in self.layers:
                    state_to_load[k] = v
                else:
                    logging.debug(
                        "Skipping parameter %s at pipeline rank %d",
                        k,
                        self.pipeline_rank,
                    )
                    skipped.add(k)
            elif any(
                k.startswith(key)
                for key in ["vision_encoder", "vision_language_adapter", "patch_merger", "pre_mm_projector_norm"]
            ):
                if self.pipeline_rank == 0:
                    state_to_load[k] = v
                else:
                    logging.debug(
                        "Skipping parameter %s at pipeline rank %d",
                        k,
                        self.pipeline_rank,
                    )
                    skipped.add(k)
            else:
                raise ValueError(f"Unexpected key {k}")
        assert set(state_dict.keys()) == skipped.union(set(state_to_load.keys()))
        super().load_state_dict(state_to_load, strict=strict, assign=assign)

    @staticmethod
    def from_folder(
        folder: Union[Path, str],
        max_batch_size: int = 1,
        num_pipeline_ranks: int = 1,
        device: Union[torch.device, str] = "cuda",
        dtype: Optional[torch.dtype] = None,
        softmax_fp32: bool = True,
    ) -> "Transformer":
        with open(Path(folder) / "params.json", "r") as f:
            model_args = TransformerArgs.from_dict(json.load(f))
        model_args.max_batch_size = max_batch_size
        if num_pipeline_ranks > 1:
            pipeline_rank = torch.distributed.get_rank()
        else:
            pipeline_rank = 0
        with torch.device("meta"):
            model = Transformer(
                model_args,
                pipeline_rank=pipeline_rank,
                num_pipeline_ranks=num_pipeline_ranks,
                softmax_fp32=softmax_fp32,
            )

        pt_model_file = Path(folder) / "consolidated.00.pth"
        safetensors_model_file = Path(folder) / "consolidated.safetensors"

        assert pt_model_file.exists() or safetensors_model_file.exists(), (
            f"Make sure either {pt_model_file} or {safetensors_model_file} exists"
        )
        assert not (pt_model_file.exists() and safetensors_model_file.exists()), (
            f"Both {pt_model_file} and {safetensors_model_file} cannot exist"
        )

        if pt_model_file.exists():
            loaded = torch.load(str(pt_model_file), mmap=True)
        else:
            loaded = safetensors.torch.load_file(str(safetensors_model_file))

        model.load_state_dict(loaded, assign=True, strict=True)

        return model.to(device=device, dtype=dtype)