vllm.model_executor.models.config ¶

class Gemma4Config(VerifyAndUpdateConfig):
    @staticmethod
    def verify_and_update_config(vllm_config: "VllmConfig") -> None:
        """Force unified attention backend for models with heterogeneous
        head dimensions.

        Some Gemma4 variants use different head dimensions for
        sliding window (head_dim) vs full attention (global_head_dim) layers.
        When global_head_dim > 256, FlashAttention rejects those layers
        (head_size <= 256 kernel limit), causing vLLM to select a different
        backend for each layer type. This mixed-backend execution produces
        numerical divergence and output corruption.

        The fix detects heterogeneous head dimensions from the model config
        and forces TRITON_ATTN (which has no head_size ceiling) for all
        layers when the user hasn't explicitly chosen a backend.

        TODO: Heterogeneous head_sizes (head_dim != global_head_dim)
        require NixlConnector changes to support per-layer KV transfer
        with different head dimensions for prefill-decode disaggregation.
        """
        hf_text_config = vllm_config.model_config.hf_text_config
        head_dim = getattr(hf_text_config, "head_dim", None)
        global_head_dim = getattr(hf_text_config, "global_head_dim", None)

        # Only force Triton when head dimensions actually differ AND the
        # larger one exceeds FlashAttention's kernel limit (head_size <= 256).
        # This avoids unnecessary backend forcing on smaller models where
        # the config carries global_head_dim but all layers can still use
        # the same FA backend.
        max_head_dim = max(head_dim or 0, global_head_dim or 0)
        if (
            head_dim is not None
            and global_head_dim is not None
            and head_dim != global_head_dim
            and max_head_dim > 256
            and vllm_config.attention_config.backend is None
        ):
            from vllm.v1.attention.backends.registry import (
                AttentionBackendEnum,
            )

            vllm_config.attention_config.backend = AttentionBackendEnum.TRITON_ATTN
            logger.info(
                "Gemma4 model has heterogeneous head dimensions "
                "(head_dim=%d, global_head_dim=%d). Forcing TRITON_ATTN "
                "backend to prevent mixed-backend numerical divergence.",
                head_dim,
                global_head_dim,
            )

class HybridAttentionMambaModelConfig(VerifyAndUpdateConfig):
    @classmethod
    def verify_and_update_config(cls, vllm_config: "VllmConfig") -> None:
        """
        Ensure that page size of attention layers is greater than or
        equal to the mamba layers. If not, automatically set the attention
        block size to ensure that it is. If the attention page size is
        strictly greater than the mamba page size, we pad the mamba page size
        to make them equal.

        Args:
            vllm_config: vLLM Config
        """
        cache_config = vllm_config.cache_config

        # Disable calculate_kv_scales for hybrid models: uninitialized
        # recurrent state corrupts scales during the calibration pass.
        # See issue: https://github.com/vllm-project/vllm/issues/37554
        if cache_config.calculate_kv_scales:
            logger.warning(
                "Disabling calculate_kv_scales for hybrid model '%s'. "
                "Hybrid models with recurrent layers (GDN, Mamba, SSM) "
                "produce unreliable KV cache scales during the "
                "calibration pass because recurrent state is "
                "uninitialized. Using default scale of 1.0 instead.",
                vllm_config.model_config.model,
            )
            cache_config.calculate_kv_scales = False

        # Save the user input before it gets modified by MambaModelConfig
        mamba_block_size = cache_config.mamba_block_size
        # Enable FULL_AND_PIECEWISE by default
        MambaModelConfig.verify_and_update_config(vllm_config)

        attention_config = vllm_config.attention_config
        cache_config = vllm_config.cache_config
        model_config = vllm_config.model_config
        parallel_config = vllm_config.parallel_config

        if cache_config.cache_dtype == "auto":
            kv_cache_dtype = model_config.dtype
        else:
            kv_cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_config.cache_dtype]

        # get attention page size (for 1 token)
        # Attention backend constraints:
        # - FlashAttention (FA) requires block size to be multiple of 16
        # - MLA (Multi-head Latent Attention) requires larger alignment:
        #   * CUTLASS_MLA backend: kernel_block_size 128 alignment
        #   * Other MLA backends: kernel_block_size 64 alignment
        if model_config.use_mla:
            use_cutlass_mla = (
                attention_config.backend == AttentionBackendEnum.CUTLASS_MLA
            )
            kernel_block_alignment_size = 128 if use_cutlass_mla else 64
            attn_page_size_1_token = MLAAttentionSpec(
                block_size=1,
                num_kv_heads=model_config.get_num_kv_heads(parallel_config),
                head_size=model_config.get_head_size(),
                dtype=kv_cache_dtype,
            ).page_size_bytes
        else:
            kernel_block_alignment_size = 16
            attn_page_size_1_token = FullAttentionSpec(
                block_size=1,
                num_kv_heads=model_config.get_num_kv_heads(parallel_config),
                head_size=model_config.get_head_size(),
                dtype=kv_cache_dtype,
            ).page_size_bytes

        model_cls, _ = ModelRegistry.resolve_model_cls(
            model_config.architecture,
            model_config=model_config,
        )

        # get mamba page size
        mamba_page_size = MambaSpec(
            shapes=model_cls.get_mamba_state_shape_from_config(vllm_config),
            dtypes=model_cls.get_mamba_state_dtype_from_config(vllm_config),
            block_size=-1,  # block_size doesn't matter for mamba page size
        ).page_size_bytes

        # Model may be marked as is_hybrid
        #  but mamba is skipped via config,
        #  return directly
        if mamba_page_size == 0:
            return

        if cache_config.mamba_cache_mode == "all":
            # With prefix caching, select attention block size to
            # optimize for mamba kernel performance

            # Mamba2 SSD kernel uses a chunk_size, e.g. 256
            # Align the block to the kernel: use lowest multiple of chunk_size
            # of attention tokens that would fit mamba_page_size:
            # e.g. for mamba page size = 788kB
            #          attn_1_token = 2kB -> fits ~394 tokens
            #      then round up to a multiple of 256 -> 512 tokens
            # End result:
            #  attn_block_size = 512
            #  mamba_block_size = 512 (aligned to a multiple of chunk_size)
            # TODO(tdoublep): this constraint can be relaxed fairly
            # easily by changing the way we layout chunks in the
            # mamba2 kernels.

            base_chunk_size = mamba_block_size or model_config.get_mamba_chunk_size()
            attn_tokens_per_mamba_state = cdiv(mamba_page_size, attn_page_size_1_token)
            chunk_size = lcm(base_chunk_size, kernel_block_alignment_size)
            attn_block_size = chunk_size * cdiv(attn_tokens_per_mamba_state, chunk_size)
            cache_config.mamba_block_size = attn_block_size
        else:
            # Without prefix caching, select minimum valid attention block size
            # to minimize mamba state padding

            # Calculate minimum attention block size that satisfies both:
            # 1. Backend alignment requirements (kernel_block_alignment_size)
            # 2. Mamba page size compatibility (attn_page_size >= mamba_page_size)
            attn_block_size = kernel_block_alignment_size * cdiv(
                mamba_page_size, kernel_block_alignment_size * attn_page_size_1_token
            )

        # override attention block size if it is too small,
        # even if the user has explicitly set it
        if cache_config.block_size < attn_block_size:
            cache_config.block_size = attn_block_size
            logger.info(
                "Setting attention block size to %d tokens "
                "to ensure that attention page size is >= mamba page size.",
                attn_block_size,
            )

        # By default, mamba block size will be set to max_model_len.
        # When enabling prefix caching and using align mamba cache
        # mode, we align mamba block size to the block size as the
        # basic granularity for prefix caching.
        if cache_config.mamba_cache_mode == "align":
            cache_config.mamba_block_size = cache_config.block_size

        # compute new attention page size
        attn_page_size = cache_config.block_size * attn_page_size_1_token

        assert attn_page_size >= mamba_page_size

        if attn_page_size == mamba_page_size:
            # don't need to pad mamba page size
            return

        # pad mamba page size to exactly match attention
        if (
            cache_config.mamba_page_size_padded is None
            or cache_config.mamba_page_size_padded != attn_page_size
        ):
            cache_config.mamba_page_size_padded = attn_page_size
            mamba_padding_pct = (
                100 * (attn_page_size - mamba_page_size) / mamba_page_size
            )
            logger.info(
                "Padding mamba page size by %.2f%% to ensure "
                "that mamba page size and attention page size are "
                "exactly equal.",
                mamba_padding_pct,
            )

class LlamaNemotronVLConfig(VerifyAndUpdateConfig):
    """Config handler for LlamaNemotronVL embedding models."""

    @staticmethod
    def verify_and_update_model_config(model_config: "ModelConfig") -> None:
        from vllm.config.pooler import SequencePoolingType

        hf_config = model_config.hf_config

        # Set bidirectional attention on the language model config
        hf_config.is_causal = False
        if hasattr(hf_config, "llm_config"):
            hf_config.llm_config.is_causal = False

        if hasattr(hf_config, "vision_config"):
            hf_config.patch_size = hf_config.vision_config.patch_size

        # Set up pooling type
        pooling_type_map: dict[str, SequencePoolingType] = {
            "avg": "MEAN",
            "cls": "CLS",
            "last": "LAST",
        }

        # Get pooling type from config (check both top-level and llm_config)
        pooling = getattr(hf_config, "pooling", None)
        if pooling is None and hasattr(hf_config, "llm_config"):
            pooling = getattr(hf_config.llm_config, "pooling", "avg")

        pooling_type = pooling_type_map.get(pooling)
        if pooling_type is None:
            raise ValueError(f"pool_type {pooling!r} not supported")

        model_config.pooler_config.seq_pooling_type = pooling_type

class MambaModelConfig(VerifyAndUpdateConfig):
    @classmethod
    def verify_and_update_config(cls, vllm_config: "VllmConfig") -> None:
        """
        Enable FULL_AND_PIECEWISE cuda graph mode by default (required
        to get good performance for mamba layers in V1).

        Args:
            vllm_config: vLLM Config
        """
        model_config = vllm_config.model_config
        cache_config = vllm_config.cache_config

        if cache_config.enable_prefix_caching:
            if cache_config.mamba_cache_mode == "none":
                cache_config.mamba_cache_mode = (
                    "all" if model_config.supports_mamba_prefix_caching else "align"
                )
                logger.warning(
                    "Mamba cache mode is set to '%s' for %s by default "
                    "when prefix caching is enabled",
                    cache_config.mamba_cache_mode,
                    model_config.architecture,
                )
            if (
                cache_config.mamba_cache_mode == "all"
                and not model_config.supports_mamba_prefix_caching
            ):
                cache_config.mamba_cache_mode = "align"
                logger.warning(
                    "Hybrid or mamba-based model detected without support "
                    "for prefix caching with Mamba cache 'all' mode: "
                    "falling back to 'align' mode."
                )
            if cache_config.mamba_cache_mode == "align":
                assert vllm_config.scheduler_config.enable_chunked_prefill, (
                    "Chunked prefill is required for mamba cache mode 'align'."
                )
            logger.info(
                "Warning: Prefix caching in Mamba cache '%s' "
                "mode is currently enabled. "
                "Its support for Mamba layers is experimental. "
                "Please report any issues you may observe.",
                cache_config.mamba_cache_mode,
            )
            # By default, mamba block size will be set to max_model_len (see
            # below). When enabling prefix caching, we align mamba block size
            # to the block size as the basic granularity for prefix caching.
            if cache_config.mamba_block_size is None:
                cache_config.mamba_block_size = cache_config.block_size
        else:
            if cache_config.mamba_cache_mode != "none":
                cache_config.mamba_cache_mode = "none"
                logger.warning(
                    "Mamba cache mode is set to 'none' when prefix caching is disabled"
                )
            if cache_config.mamba_block_size is None:
                cache_config.mamba_block_size = model_config.max_model_len

class NemotronHForCausalLMConfig(VerifyAndUpdateConfig):
    @staticmethod
    def verify_and_update_config(vllm_config: "VllmConfig") -> None:
        """Update mamba_ssm_cache_dtype for NemotronH models when set to 'auto'
        (or not explicitly set), to the value specified in the HF config, or to
        float16 if not specified.
        """
        cache_config = vllm_config.cache_config
        if cache_config.mamba_ssm_cache_dtype == "auto":
            hf_config = vllm_config.model_config.hf_config
            mamba_ssm_cache_dtype = getattr(
                hf_config, "mamba_ssm_cache_dtype", "float16"
            )
            logger.info(
                "Updating mamba_ssm_cache_dtype to '%s' for NemotronH model",
                mamba_ssm_cache_dtype,
            )
            cache_config.mamba_ssm_cache_dtype = mamba_ssm_cache_dtype

class Qwen3_5ForConditionalGenerationConfig(VerifyAndUpdateConfig):
    @staticmethod
    def verify_and_update_config(vllm_config: "VllmConfig") -> None:
        """Update mamba_ssm_cache_dtype for Qwen3.5 models when set to 'auto'
        (or not explicitly set), to the value specified in the HF config's
        mamba_ssm_dtype field. Warn if the user explicitly overrides it to a
        different value.
        """
        cache_config = vllm_config.cache_config
        hf_text_config = vllm_config.model_config.hf_text_config
        mamba_ssm_dtype = getattr(hf_text_config, "mamba_ssm_dtype", None)
        if cache_config.mamba_ssm_cache_dtype == "auto":
            if mamba_ssm_dtype is not None:
                cache_config.mamba_ssm_cache_dtype = mamba_ssm_dtype
        elif (
            mamba_ssm_dtype is not None
            and cache_config.mamba_ssm_cache_dtype != mamba_ssm_dtype
        ):
            logger.warning(
                "Qwen3.5 model specifies mamba_ssm_dtype='%s' in its config, "
                "but --mamba-ssm-cache-dtype='%s' was passed. "
                "Using the user-specified value.",
                mamba_ssm_dtype,
                cache_config.mamba_ssm_cache_dtype,
            )