vllm.model_executor.models.sarvam ¶

class SarvamMLAForCausalLM(
    nn.Module, SupportsPP, SupportsLoRA, SupportsEagle3, SarvamMixtureOfExperts
):
    packed_modules_mapping = {
        "q_proj": ["q_proj"],
        "q_a_proj": ["q_a_proj"],
        "q_b_proj": ["q_b_proj"],
        "kv_a_proj_with_mqa": ["kv_a_proj_with_mqa"],
        "kv_b_proj": ["kv_b_proj"],
        "gate_up_proj": ["gate_proj", "up_proj"],
    }

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        self.config = config
        self.quant_config = quant_config

        self.model = SarvamMLAModel(
            vllm_config=vllm_config,
            prefix=maybe_prefix(prefix, "model"),
        )

        self.tie_word_embeddings = getattr(config, "tie_word_embeddings", False)
        if get_pp_group().is_last_rank:
            self.lm_head = ParallelLMHead(
                config.vocab_size,
                config.hidden_size,
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "lm_head"),
            )
            if self.tie_word_embeddings:
                self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
            self.logits_processor = LogitsProcessor(config.vocab_size)
        else:
            self.lm_head = PPMissingLayer()
            self.logits_processor = None  # type: ignore

        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors
        )

        self.num_moe_layers = 0

        self.moe_layers = []
        self.moe_mlp_layers = []

        example_moe = None
        for layer in self.model.layers:
            if isinstance(layer, PPMissingLayer):
                continue
            if isinstance(layer.mlp, SarvamMLAMoE):
                example_moe = layer.mlp
                self.moe_mlp_layers.append(layer.mlp)
                self.moe_layers.append(layer.mlp.experts)
                self.num_moe_layers += 1

        self.extract_moe_parameters(example_moe)

    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]:
        """Return backbone outputs, including auxiliary states when captured."""
        return self.model(
            input_ids=input_ids,
            positions=positions,
            intermediate_tensors=intermediate_tensors,
            inputs_embeds=inputs_embeds,
        )

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor | None:
        if not get_pp_group().is_last_rank:
            return None
        logits = self.logits_processor(self.lm_head, hidden_states)
        return logits

    def load_weights(
        self,
        weights: Iterable[tuple[str, torch.Tensor]],
    ) -> set[str]:
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights)

@support_torch_compile(
    dynamic_arg_dims={
        "input_ids": 0,
        "positions": 0,
        "intermediate_tensors": 0,
        "inputs_embeds": 0,
    }
)
class SarvamMLAModel(nn.Module, EagleModelMixin):
    """Sarvam MLA backbone with stage-local EAGLE3 auxiliary capture."""

    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_stacked={
            # .experts.gate_up_proj must be handled by MoERunner.load_weights for EP
            ".mlp.gate_proj": (".mlp.gate_up_proj", 0),
            ".mlp.up_proj": (".mlp.gate_up_proj", 1),
            ".shared_experts.gate_proj": (".shared_experts.gate_up_proj", 0),
            ".shared_experts.up_proj": (".shared_experts.gate_up_proj", 1),
        }
    )

    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
    ) -> None:
        super().__init__()

        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config

        self.config = config
        self.vocab_size = config.vocab_size
        self.embed_dim = config.hidden_size
        self.tie_word_embeddings = getattr(config, "tie_word_embeddings", False)
        if get_pp_group().is_first_rank or (
            self.tie_word_embeddings and get_pp_group().is_last_rank
        ):
            self.embed_tokens = VocabParallelEmbedding(
                self.vocab_size,
                self.embed_dim,
                quant_config=quant_config,
                prefix=f"{prefix}.embed_tokens",
            )
        else:
            self.embed_tokens = PPMissingLayer()

        self.embedding_dropout = torch.nn.Dropout(
            getattr(config, "embedding_dropout", 0.0)
        )
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: SarvamMLABlock(
                vllm_config=vllm_config,
                prefix=prefix,
            ),
            prefix=f"{prefix}.layers",
        )
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size
        )
        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(self.embed_dim, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()

    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]:
        """Run this stage and optionally return its auxiliary hidden states.

        Auxiliary captures are local to this stage, matching Qwen3 MoE.
        Pipeline transport carries only hidden states and residual; EAGLE3
        capture across pipeline stages is not supported by this model.

        Returns:
            Intermediate tensors on non-final stages. On the final stage,
            normalized hidden states, paired with auxiliary states if captured.
        """
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.embed_input_ids(input_ids)
            hidden_states = self.embedding_dropout(hidden_states)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

        aux_hidden_states = self._maybe_add_hidden_state(
            [], self.start_layer, hidden_states, residual
        )
        for layer_idx, layer in enumerate(
            islice(self.layers, self.start_layer, self.end_layer),
            start=self.start_layer,
        ):
            hidden_states, residual = layer(
                hidden_states,
                positions,
                residual,
            )
            self._maybe_add_hidden_state(
                aux_hidden_states, layer_idx + 1, hidden_states, residual
            )

        if not get_pp_group().is_last_rank:
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
        if residual is None:
            hidden_states = self.norm(hidden_states)
        else:
            hidden_states, _ = self.norm(hidden_states, residual)

        if len(aux_hidden_states) > 0:
            return hidden_states, aux_hidden_states
        return hidden_states

    def load_weights(
        self,
        weights: Iterable[tuple[str, torch.Tensor]],
    ) -> set[str]:
        loader = AutoWeightsLoader(self)
        return loader.load_weights(
            _normalized_weights(weights), mapper=self.hf_to_vllm_mapper
        )

class SarvamMoEForCausalLM(BailingMoeForCausalLM):
    """Same as BailingMoeForCausalLM, but normalizes gate expert_bias pre-load."""

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        return super().load_weights(_normalized_weights(weights))