# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project from copy import deepcopy from typing import TYPE_CHECKING import vllm.envs as envs from vllm.logger import init_logger from vllm.model_executor.models import ModelRegistry from vllm.utils import cdiv, round_up from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE from vllm.v1.kv_cache_interface import FullAttentionSpec, MambaSpec if TYPE_CHECKING: from vllm.config import VllmConfig logger = init_logger(__name__) class VerifyAndUpdateConfig: @staticmethod def verify_and_update_config(vllm_config: "VllmConfig") -> None: raise NotImplementedError class Gemma3TextModelConfig: @staticmethod def verify_and_update_config(vllm_config: "VllmConfig") -> None: hf_config = vllm_config.model_config.hf_config hf_config.is_causal = not hf_config.use_bidirectional_attention class GteNewModelConfig(VerifyAndUpdateConfig): @staticmethod def verify_and_update_config(vllm_config: "VllmConfig") -> None: config = vllm_config.model_config.hf_config assert config.__class__.__name__ == "NewConfig" assert config.hidden_act == "gelu" config.hidden_act = "geglu" head_dim = config.hidden_size // config.num_attention_heads config.rotary_kwargs = { "head_size": head_dim, "rotary_dim": getattr(config, "rotary_emb_dim", head_dim), "max_position": config.max_position_embeddings, "base": config.rope_theta, "rope_scaling": getattr(config, "rope_scaling", None), } class JambaForSequenceClassificationConfig(VerifyAndUpdateConfig): @staticmethod def verify_and_update_config(vllm_config: "VllmConfig") -> None: pooler_config = vllm_config.model_config.pooler_config if pooler_config.activation is None: pooler_config.activation = False class JinaRobertaModelConfig(VerifyAndUpdateConfig): @staticmethod def verify_and_update_config(vllm_config: "VllmConfig") -> None: model_config = vllm_config.model_config config = model_config.hf_config if config.position_embedding_type == "rotary": assert config.__class__.__name__ == "XLMRobertaFlashConfig" head_dim = config.hidden_size // config.num_attention_heads max_position = config.max_position_embeddings # Jina-embeddings-v3 has max_position_embeddings=8194, which will cause # out-of-bound index issue at RoPE for long prompts with torch.compile, # because it can't be divided by triton num_warps(default=4 or 8). # To deal with this, we increase max_position to multiple of n_warps, # so that triton kernel won't hit out-of-bound index in RoPE cache. if not model_config.enforce_eager: max_position = round_up(max_position, 8) config.rotary_kwargs = { "head_size": head_dim, "rotary_dim": getattr(config, "rotary_emb_dim", head_dim), "max_position": max_position, "base": getattr(config, "rope_theta", config.rotary_emb_base), "rope_scaling": getattr(config, "rope_scaling", None), } class NomicBertModelConfig(VerifyAndUpdateConfig): @staticmethod def verify_and_update_config(vllm_config: "VllmConfig") -> None: config = vllm_config.model_config.hf_config assert config.__class__.__name__ == "NomicBertConfig" assert config.activation_function in ["swiglu", "gelu"] config.position_embedding_type = getattr( config, "position_embedding_type", "rope" ) if config.activation_function == "swiglu": config.hidden_act = "silu" else: config.hidden_act = config.activation_function assert config.mlp_fc1_bias == config.mlp_fc2_bias == config.qkv_proj_bias config.bias = config.qkv_proj_bias assert config.rotary_emb_scale_base is None assert not config.rotary_emb_interleaved config.layer_norm_eps = config.layer_norm_epsilon config.intermediate_size = config.n_inner config.hidden_size = config.n_embd config.num_hidden_layers = config.n_layer head_dim = config.hidden_size // config.num_attention_heads rotary_emb_dim = int(head_dim * config.rotary_emb_fraction) max_trained_positions = getattr(config, "max_trained_positions", 2048) config.rotary_kwargs = { "head_size": head_dim, "rotary_dim": rotary_emb_dim, "max_position": max_trained_positions, "base": getattr(config, "rope_theta", config.rotary_emb_base), "rope_scaling": getattr(config, "rope_scaling", None), } # we ignore config.rotary_scaling_factor so that for datasets shorter # than max_trained_positions 2048, the results are consistent # with SentenceTransformer. # The context extension uses vllm style rope_theta and rope_scaling. # See #17785 #18755 if ( not vllm_config.model_config.hf_overrides and vllm_config.model_config.original_max_model_len is None ): # Default # Reset max_model_len to max_trained_positions. # nomic-embed-text-v2-moe the length is set to 512 # by sentence_bert_config.json. max_model_len_before = vllm_config.model_config.max_model_len max_model_len = min( vllm_config.model_config.max_model_len, max_trained_positions ) vllm_config.recalculate_max_model_len(max_model_len) logger.warning( "Nomic context extension is disabled. " "Changing max_model_len from %s to %s. " "To enable context extension, see: " "https://github.com/vllm-project/vllm/tree/main/examples/offline_inference/context_extension.html", max_model_len_before, vllm_config.model_config.max_model_len, ) else: # We need to re-verify max_model_len to avoid lengths # greater than position_embedding. model_config = vllm_config.model_config hf_text_config = model_config.hf_text_config if isinstance(model_config.hf_overrides, dict): # hf_overrides_kw max_model_len = model_config.hf_overrides.get( "max_model_len", vllm_config.model_config.max_model_len ) else: # hf_overrides_fn # This might be overridden by sentence_bert_config.json. max_model_len = vllm_config.model_config.max_model_len # reset hf_text_config for recalculate_max_model_len. if hasattr(hf_text_config, "max_model_len"): delattr(hf_text_config, "max_model_len") hf_text_config.max_position_embeddings = max_trained_positions hf_text_config.rope_scaling = config.rotary_kwargs["rope_scaling"] # The priority of sentence_bert_config.json is higher # than max_position_embeddings encoder_config = deepcopy(model_config.encoder_config) encoder_config.pop("max_seq_length", None) model_config.encoder_config = encoder_config vllm_config.recalculate_max_model_len(max_model_len) class Qwen2ForProcessRewardModelConfig(VerifyAndUpdateConfig): @staticmethod def verify_and_update_config(vllm_config: "VllmConfig") -> None: pooler_config = vllm_config.model_config.pooler_config if pooler_config.step_tag_id is None: pooler_config.step_tag_id = 151651 class Qwen2ForRewardModelConfig(VerifyAndUpdateConfig): @staticmethod def verify_and_update_config(vllm_config: "VllmConfig") -> None: pooler_config = vllm_config.model_config.pooler_config if pooler_config.softmax is None: pooler_config.softmax = False class Qwen3ForSequenceClassificationConfig(VerifyAndUpdateConfig): @staticmethod def verify_and_update_config(vllm_config: "VllmConfig") -> None: config = vllm_config.model_config.hf_config is_original_qwen3_reranker = getattr( config, "is_original_qwen3_reranker", False ) if not is_original_qwen3_reranker: return tokens = getattr(config, "classifier_from_token", None) assert tokens is not None and len(tokens) == 2, ( "Try loading the original Qwen3 Reranker?, see: " "https://github.com/vllm-project/vllm/tree/main/examples/offline_inference/qwen3_reranker.py" ) vllm_config.model_config.hf_config.method = "from_2_way_softmax" class JinaVLForSequenceClassificationConfig(VerifyAndUpdateConfig): @staticmethod def verify_and_update_config(vllm_config: "VllmConfig") -> None: config = vllm_config.model_config.hf_config config.num_labels = 1 pooler_config = vllm_config.model_config.pooler_config if pooler_config.logit_bias is None: pooler_config.logit_bias = 2.65 class SnowflakeGteNewModelConfig(VerifyAndUpdateConfig): @staticmethod def verify_and_update_config(vllm_config: "VllmConfig") -> None: config = vllm_config.model_config.hf_config assert config.__class__.__name__ == "GteConfig" assert config.hidden_act == "gelu" config.hidden_act = "geglu" head_dim = config.hidden_size // config.num_attention_heads config.rotary_kwargs = { "head_size": head_dim, "rotary_dim": getattr(config, "rotary_emb_dim", head_dim), "max_position": config.max_position_embeddings, "base": config.rope_theta, "rope_scaling": getattr(config, "rope_scaling", None), } class GptOssForCausalLMConfig(VerifyAndUpdateConfig): @staticmethod def verify_and_update_config(vllm_config: "VllmConfig") -> None: structured_outputs_config = vllm_config.structured_outputs_config if structured_outputs_config.reasoning_parser == "": structured_outputs_config.reasoning_parser = "openai_gptoss" # Increase the max capture size from 512 to 992 for performance. # NOTE(woosuk): This will increase the number of CUDA graphs # from 67 to 81. scheduler_config = vllm_config.scheduler_config if len(scheduler_config.cuda_graph_sizes) == 1: max_capture_size = scheduler_config.cuda_graph_sizes[0] # FIXME(woosuk): When using full cuda graph with FA3, the max # supported size is 992. if max_capture_size < 992: cuda_graph_sizes = [1, 2, 4] # Step size 8 for small batch sizes cuda_graph_sizes += [i for i in range(8, 256, 8)] # Step size 16 for larger batch sizes cuda_graph_sizes += [i for i in range(256, 993, 16)] scheduler_config.cuda_graph_sizes = cuda_graph_sizes logger.info( "Overriding max cuda graph capture size to %d for performance.", 992 ) 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 """ if not envs.VLLM_USE_V1: return model_config = vllm_config.model_config cache_config = vllm_config.cache_config # Set mamba block size to max_model_len (this may get # override by prefix caching logic later) cache_config.mamba_block_size = model_config.max_model_len # TODO(@tdoublep) find a better way to do this than whitelist MAMBA2_MODELS = [ "BambaForCausalLM", "FalconH1ForCausalLM", "GraniteMoeHybridForCausalLM", "Mamba2ForCausalLM", "NemotronHForCausalLM", "Zamba2ForCausalLM", ] if cache_config.enable_prefix_caching: if model_config.architecture in MAMBA2_MODELS: logger.info( "Warning: Prefix caching is currently enabled. " "Its support for Mamba2 layers is experimental. " "Please report any issues you may observe." ) else: logger.info( "Hybrid or mamba-based model detected without " "support for prefix caching: disabling." ) cache_config.enable_prefix_caching = False # TODO(tdoublep): remove once cascade attention is supported logger.info( "Disabling cascade attention since it is not supported for hybrid models." ) model_config.disable_cascade_attn = True 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 """ if not envs.VLLM_USE_V1: return # Enable FULL_AND_PIECEWISE by default MambaModelConfig.verify_and_update_config(vllm_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) 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=model_config.max_model_len, ).page_size_bytes # Model may be marked as is_hybrid # but mamba is skipped via config, # return directly if mamba_page_size == 0: return # Attention backend constraints: # - FlashAttention (FA) requires block size to be multiple of 16 # - MLA (Multi-head Latent Attention) requires larger alignment: # * CUTLASS_MLA backend: 128-byte alignment # * Other MLA backends: 64-byte alignment if model_config.use_mla: use_cutlass_mla = envs.VLLM_ATTENTION_BACKEND == "CUTLASS_MLA" kernel_block_alignment_size = 128 if use_cutlass_mla else 64 else: kernel_block_alignment_size = 16 if cache_config.enable_prefix_caching: # With prefix caching, select attention block size to # optimize for mamba kernel performance # mamba 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 mulitple 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. from math import gcd def lcm(a, b): return a * b // gcd(a, b) base_chunk_size = 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 either (a) the # user has not set it or (b) the user has set it # too small. if cache_config.block_size is None or 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, ) # 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 DeepseekV32ForCausalLM(VerifyAndUpdateConfig): @classmethod def verify_and_update_config(cls, vllm_config: "VllmConfig") -> None: """ Updated fp8 cache to custom "fp8_ds_mla" format for DeepSeekV32 """ hf_config = vllm_config.model_config.hf_config # Mirror the check in vllm/model_executor/models/deepseek_v2.py is_v32 = hasattr(hf_config, "index_topk") assert is_v32 # For DeepSeekV3.2, a custom fp8 format is used when fp8 kv-cache is enabled. cache_config = vllm_config.cache_config if cache_config.cache_dtype.startswith("fp8"): cache_config.cache_dtype = "fp8_ds_mla" logger.info("Using custom fp8 kv-cache format for DeepSeekV3.2") if cache_config.cache_dtype == "bfloat16": cache_config.cache_dtype = "auto" logger.info("Using bfloat16 kv-cache for DeepSeekV3.2") MODELS_CONFIG_MAP: dict[str, type[VerifyAndUpdateConfig]] = { "GteModel": SnowflakeGteNewModelConfig, "GteNewModel": GteNewModelConfig, "GteNewForSequenceClassification": GteNewModelConfig, "Gemma3TextModel": Gemma3TextModelConfig, "NomicBertModel": NomicBertModelConfig, "Qwen2ForProcessRewardModel": Qwen2ForProcessRewardModelConfig, "Qwen2ForRewardModel": Qwen2ForRewardModelConfig, "Qwen3ForSequenceClassification": Qwen3ForSequenceClassificationConfig, "XLMRobertaModel": JinaRobertaModelConfig, "JinaVLForRanking": JinaVLForSequenceClassificationConfig, "JambaForSequenceClassification": JambaForSequenceClassificationConfig, "GptOssForCausalLM": GptOssForCausalLMConfig, "MambaForCausalLM": MambaModelConfig, "Mamba2ForCausalLM": MambaModelConfig, "FalconMambaForCausalLM": MambaModelConfig, "DeepseekV32ForCausalLM": DeepseekV32ForCausalLM, }