vllm.model_executor.models.chameleon ¶

class ChameleonAttention(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[dict[str, Any]] = None,
        max_position_embeddings: int = 4096,
        quant_config: Optional[QuantizationConfig] = None,
        bias: bool = False,
        cache_config: Optional[CacheConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = num_kv_heads
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        self.head_dim = hidden_size // self.total_num_heads
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        self.qkv_proj = QKVParallelLinear(
            hidden_size=hidden_size,
            head_size=self.head_dim,
            total_num_heads=self.total_num_heads,
            total_num_kv_heads=self.total_num_kv_heads,
            bias=bias,
            quant_config=quant_config,
        )
        self.o_proj = RowParallelLinear(
            input_size=self.total_num_heads * self.head_dim,
            output_size=hidden_size,
            bias=bias,
            quant_config=quant_config,
        )
        self.q_norm = ChameleonLayerNorm((self.num_heads, self.head_dim))
        self.k_norm = ChameleonLayerNorm((self.num_kv_heads, self.head_dim))
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            rope_scaling=rope_scaling,
        )

        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
        )

    def _apply_qk_norm(
        self, q: torch.Tensor, k: torch.Tensor
    ) -> tuple[torch.Tensor, torch.Tensor]:
        # reshape for layernorm
        q = q.reshape(-1, self.num_heads, self.head_dim)
        k = k.reshape(-1, self.num_kv_heads, self.head_dim)
        q = self.q_norm(q)
        k = self.k_norm(k)
        q = q.view(*q.shape[:-2], -1)
        k = k.view(*k.shape[:-2], -1)
        return q, k

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q, k = self._apply_qk_norm(q, k)

        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.o_proj(attn_output)
        return output

class ChameleonDecoderLayer(nn.Module):
    def __init__(
        self,
        config: ChameleonConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        if rope_scaling is not None and getattr(
            config, "original_max_position_embeddings", None
        ):
            rope_scaling["original_max_position_embeddings"] = (
                config.original_max_position_embeddings
            )
        max_position_embeddings = getattr(config, "max_position_embeddings", 4096)

        self.self_attn = ChameleonAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=getattr(
                config, "num_key_value_heads", config.num_attention_heads
            ),
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            quant_config=quant_config,
            bias=False,
            cache_config=cache_config,
            prefix=f"{prefix}.self_attn",
        )
        self.mlp = ChameleonMLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
            bias=getattr(config, "mlp_bias", False),
        )
        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(
            config.hidden_size, eps=config.rms_norm_eps
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
            hidden_states, residual = self.input_layernorm(hidden_states, residual)
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )

        # Fully Connected
        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
        hidden_states = self.mlp(hidden_states)

        return hidden_states, residual

class ChameleonDummyInputsBuilder(BaseDummyInputsBuilder[ChameleonProcessingInfo]):
    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
        num_images = mm_counts.get("image", 0)

        processor = self.info.get_hf_processor()
        image_token = processor.image_token

        return image_token * num_images

    def get_dummy_mm_data(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
        mm_options: Optional[Mapping[str, BaseDummyOptions]] = None,
    ) -> MultiModalDataDict:
        config = self.info.get_hf_config()

        width = height = config.vq_config.resolution
        num_images = mm_counts.get("image", 0)

        image_overrides = mm_options.get("image") if mm_options else None

        return {
            "image": self._get_dummy_images(
                width=width,
                height=height,
                num_images=num_images,
                overrides=image_overrides,
            )
        }

@MULTIMODAL_REGISTRY.register_processor(
    ChameleonMultiModalProcessor,
    info=ChameleonProcessingInfo,
    dummy_inputs=ChameleonDummyInputsBuilder,
)
class ChameleonForConditionalGeneration(
    nn.Module, SupportsMultiModal, SupportsPP, SupportsQuant
):
    merge_by_field_config = True

    packed_modules_mapping = {
        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
        "gate_up_proj": ["gate_proj", "up_proj"],
    }

    @classmethod
    def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
        if modality.startswith("image"):
            return "<image>"

        raise ValueError("Only image modality is supported")

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        multimodal_config = vllm_config.model_config.multimodal_config
        self.config = config
        self.multimodal_config = multimodal_config
        self.model = ChameleonModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
        self.unpadded_vocab_size = config.vocab_size
        self.lm_head = ParallelLMHead(
            self.unpadded_vocab_size,
            config.hidden_size,
            prefix=maybe_prefix(prefix, "lm_head"),
        )
        if config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight

        logit_scale = getattr(config, "logit_scale", 1.0)
        self.logits_processor = LogitsProcessor(
            self.unpadded_vocab_size, config.vocab_size, logit_scale
        )
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors
        )

    def _parse_and_validate_image_input(
        self, **kwargs: object
    ) -> Optional[ChameleonImagePixelInputs]:
        pixel_values = kwargs.pop("pixel_values", None)

        if pixel_values is None:
            return None

        vq_config: ChameleonVQVAEConfig = self.config.vq_config
        expected_h = expected_w = vq_config.resolution

        return ChameleonImagePixelInputs(
            type="pixel_values",
            data=pixel_values,
            resolve_bindings={"h": expected_h, "w": expected_w},
        )

    def get_language_model(self) -> torch.nn.Module:
        return self.model

    def get_multimodal_embeddings(self, **kwargs: object) -> MultiModalEmbeddings:
        image_input = self._parse_and_validate_image_input(**kwargs)
        if image_input is None:
            return []
        assert self.model.vqmodel is not None
        image_tokens = self.model.get_image_tokens(
            image_input["data"].to(self.config.torch_dtype)
        )
        vision_embeddings = self.model.get_input_embeddings(image_tokens)
        return vision_embeddings

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        if intermediate_tensors is not None:
            inputs_embeds = None

        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds=inputs_embeds
        )
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> Optional[torch.Tensor]:
        logits = self.logits_processor(self.lm_head, hidden_states)

        # Disallow image tokens which does not include special
        # begin-image and end-image tokens
        if logits is not None:
            image_tokens = self.model.vocabulary_mapping.image_tokens
            logits[:, image_tokens] = torch.finfo(logits.dtype).min

        return logits

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
            (".gate_up_proj", ".gate_proj", 0),
            (".gate_up_proj", ".up_proj", 1),
        ]
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue

            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
                # Models trained using ColossalAI may include these tensors in
                # the checkpoint. Skip them.
                continue

            # With tie_word_embeddings, we can skip lm_head.weight
            # The weight might appear unnecessarily in the files if the model is
            # processed with quantization, LoRA, fine-tuning, etc.
            if self.config.tie_word_embeddings and "lm_head.weight" in name:
                continue

            use_default_weight_loading = False
            if "vqmodel" in name:
                if self.model.vqmodel is not None:
                    # We only do sharding for language model and
                    # not vqvae for now.
                    use_default_weight_loading = True
            else:
                for param_name, weight_name, shard_id in stacked_params_mapping:
                    if weight_name not in name:
                        continue
                    name = name.replace(weight_name, param_name)
                    # Skip loading extra bias for GPTQ models.
                    if name.endswith(".bias") and name not in params_dict:
                        continue
                    if is_pp_missing_parameter(name, self):
                        continue
                    param = params_dict[name]
                    weight_loader = param.weight_loader
                    weight_loader(param, loaded_weight, shard_id)
                    break
                else:
                    # Skip loading extra bias for GPTQ models.
                    if name.endswith(".bias") and name not in params_dict:
                        continue
                    # Remapping the name of FP8 kv-scale.
                    if name.endswith("kv_scale"):
                        remapped_kv_scale_name = name.replace(
                            ".kv_scale", ".attn.kv_scale"
                        )
                        if remapped_kv_scale_name not in params_dict:
                            logger.warning_once(
                                "Found kv scale in the checkpoint (e.g. %s), but not found the expected name in the model (e.g. %s). kv-scale is not loaded.",  # noqa: E501
                                name,
                                remapped_kv_scale_name,
                            )
                            continue
                        else:
                            name = remapped_kv_scale_name
                    if is_pp_missing_parameter(name, self):
                        continue
                    param = params_dict[name]
                    weight_loader = getattr(
                        param, "weight_loader", default_weight_loader
                    )
                    weight_loader(param, loaded_weight)
            if use_default_weight_loading and name in params_dict:
                if is_pp_missing_parameter(name, self):
                    continue
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params