vllm.model_executor.models.interns1_pro ¶

@MULTIMODAL_REGISTRY.register_processor(
    Qwen3VLMultiModalProcessor,
    info=InternS1ProProcessingInfo,
    dummy_inputs=Qwen3VLDummyInputsBuilder,
)
class InternS1ProForConditionalGeneration(
    Qwen3VLForConditionalGeneration, Qwen3VLMoeMixtureOfExperts
):
    is_3d_moe_weight: bool = True
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
    }

    # To ensure correct weight loading and mapping.
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={
            "model.visual.": "visual.",
            "lm_head.": "language_model.lm_head.",
            "model.language_model.": "language_model.model.",
        },
        orig_to_new_suffix={
            # Handle FOPE rotary embeddings
            ".rotary_emb.sin_coef": ".layers.0.self_attn.rotary_emb.sin_coef",
            ".rotary_emb.cos_coef": ".layers.0.self_attn.rotary_emb.cos_coef",
        },
    )

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config: PretrainedConfig = vllm_config.model_config.hf_config
        multimodal_config = vllm_config.model_config.multimodal_config

        self.config = config
        self.multimodal_config = multimodal_config
        self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
        self.video_pruning_rate = multimodal_config.video_pruning_rate
        self.is_multimodal_pruning_enabled = (
            multimodal_config.is_multimodal_pruning_enabled()
        )

        if not multimodal_config.get_limit_per_prompt(
            "image"
        ) and not multimodal_config.get_limit_per_prompt("video"):
            self.visual = None
        else:
            self.visual = Qwen3_VisionTransformer(
                config.vision_config,
                norm_eps=getattr(config, "rms_norm_eps", 1e-6),
                multimodal_config=multimodal_config,
                prefix=maybe_prefix(prefix, "visual"),
            )

        self.language_model = InternS1ProMoeLLMForCausalLM(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "language_model")
        )
        # Whether to include the gate_up_proj mapping is determined by
        # the language model.
        self.packed_modules_mapping = (
            self.packed_modules_mapping | self.language_model.packed_modules_mapping
        )

        self.make_empty_intermediate_tensors = (
            self.language_model.make_empty_intermediate_tensors
        )

        self.use_deepstack = hasattr(config.vision_config, "deepstack_visual_indexes")
        self.deepstack_num_level = (
            len(config.vision_config.deepstack_visual_indexes)
            if self.use_deepstack
            else 0
        )
        self.visual_dim = config.vision_config.out_hidden_size
        self.multiscale_dim = self.visual_dim * self.deepstack_num_level

        # Set MoE hyperparameters
        self.set_moe_parameters()

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
        """load weights"""
        skip_prefixes = ["model.time_series."]
        if self.visual is None:
            skip_prefixes.append("visual.")
        loader = AutoWeightsLoader(self, skip_prefixes=skip_prefixes)
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)

class InternS1ProMoeAttention(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_parameters: dict[str, Any],
        max_position_embeddings: int = 32768,
        head_dim: int | None = None,
        rms_norm_eps: float = 1e-06,
        qkv_bias: bool = False,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
        dual_chunk_attention_config: dict[str, Any] | None = None,
    ) -> 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 = head_dim or (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.max_position_embeddings = max_position_embeddings
        self.dual_chunk_attention_config = dual_chunk_attention_config

        self.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=qkv_bias,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )

        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )

        rope_parameters["num_key_value_heads"] = self.num_kv_heads
        self.rotary_emb = get_rope(
            self.head_dim,
            max_position=max_position_embeddings,
            rope_parameters=rope_parameters,
            dual_chunk_attention_config=dual_chunk_attention_config,
        )

        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",
            **{
                "layer_idx": extract_layer_index(prefix),
                "dual_chunk_attention_config": dual_chunk_attention_config,
            }
            if dual_chunk_attention_config
            else {},
        )

        self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
        self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)

    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)
        # Add qk-norm
        q_by_head = q.view(*q.shape[:-1], q.shape[-1] // self.head_dim, self.head_dim)
        q_by_head = self.q_norm(q_by_head)
        q = q_by_head.view(q.shape)

        k_by_head = k.view(*k.shape[:-1], k.shape[-1] // self.head_dim, self.head_dim)
        k_by_head = self.k_norm(k_by_head)
        k = k_by_head.view(k.shape)
        q, k = self.rotary_emb.forward_native(positions, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.o_proj(attn_output)
        return output

class InternS1ProMoeDecoderLayer(nn.Module):
    def __init__(self, vllm_config: VllmConfig, prefix: str = "") -> None:
        super().__init__()

        config = vllm_config.model_config.hf_text_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config

        self.hidden_size = config.hidden_size
        max_position_embeddings = getattr(config, "max_position_embeddings", 32768)
        dual_chunk_attention_config = getattr(
            config, "dual_chunk_attention_config", None
        )

        # update rope related parameters
        rope_scaling = config.rope_scaling
        fope_keys = {"fope_init_factor", "fope_sep_head", "num_inv_freq"}
        use_fope = any(rope_scaling.get(key) is not None for key in fope_keys)
        fope_init_factor = rope_scaling.get("fope_init_factor", None)
        fope_sep_head = rope_scaling.get("fope_sep_head", None)
        num_inv_freq = rope_scaling.get("num_inv_freq", None)

        config.rope_parameters["use_fope"] = use_fope
        config.rope_parameters["fope_init_factor"] = fope_init_factor
        config.rope_parameters["fope_sep_head"] = fope_sep_head
        config.rope_parameters["num_inv_freq"] = num_inv_freq

        assert use_fope, "should use FOPE for InternS1Pro model"
        self.self_attn = InternS1ProMoeAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            rope_parameters=config.rope_parameters,
            max_position_embeddings=max_position_embeddings,
            rms_norm_eps=config.rms_norm_eps,
            qkv_bias=getattr(config, "attention_bias", False),
            head_dim=getattr(config, "head_dim", None),
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
            dual_chunk_attention_config=dual_chunk_attention_config,
        )

        # `mlp_only_layers` in the config.
        layer_idx = extract_layer_index(prefix)
        mlp_only_layers = (
            [] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers
        )
        if (layer_idx not in mlp_only_layers) and (
            config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
        ):
            self.mlp = InternS1ProMoeSparseMoeBlock(
                vllm_config=vllm_config, prefix=f"{prefix}.mlp"
            )
        else:
            self.mlp = InternS1ProMoeMLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
            )
        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: torch.Tensor | None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        # Self Attention
        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 InternS1ProMoeLLMForCausalLM(Qwen3MoeForCausalLM):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        self.config = vllm_config.model_config.hf_config.text_config
        self.quant_config = vllm_config.quant_config
        self.model = InternS1ProMoeLLMModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
        self.lm_head = ParallelLMHead(
            self.config.vocab_size,
            self.config.hidden_size,
            quant_config=self.quant_config,
            prefix=maybe_prefix(prefix, "lm_head"),
        )
        if self.config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
        self.logits_processor = LogitsProcessor(self.config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors
        )

class InternS1ProMoeLLMModel(Qwen3MoeLLMModel):
    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
        decoder_layer_type: type[torch.nn.Module] = InternS1ProMoeDecoderLayer,
    ):
        super().__init__(
            vllm_config=vllm_config,
            prefix=prefix,
            decoder_layer_type=decoder_layer_type,
        )

class InternS1ProMoeMLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: QuantizationConfig | None = None,
        reduce_results: bool = True,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size,
            [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            reduce_results=reduce_results,
            prefix=f"{prefix}.down_proj",
        )
        if hidden_act != "silu":
            raise ValueError(
                f"Unsupported activation: {hidden_act}. Only silu is supported for now."
            )
        self.act_fn = SiluAndMul()

    def forward(self, x):
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x