From 65e655d2959111d508ad97515c85be0627a7b916 Mon Sep 17 00:00:00 2001 From: Woosuk Kwon Date: Fri, 26 Jun 2026 14:09:05 -0700 Subject: [PATCH] [GLM-5] Add DSV3.2/GLM5 to `vllm/models/` (#46808) Signed-off-by: Woosuk Kwon --- vllm/models/deepseek_v32/__init__.py | 22 + vllm/models/deepseek_v32/nvidia/__init__.py | 2 + vllm/models/deepseek_v32/nvidia/attention.py | 423 +++++++++++++++++++ vllm/models/deepseek_v32/nvidia/model.py | 333 +++++++++++++++ vllm/models/deepseek_v32/nvidia/mtp.py | 390 +++++++++++++++++ 5 files changed, 1170 insertions(+) create mode 100644 vllm/models/deepseek_v32/__init__.py create mode 100644 vllm/models/deepseek_v32/nvidia/__init__.py create mode 100644 vllm/models/deepseek_v32/nvidia/attention.py create mode 100644 vllm/models/deepseek_v32/nvidia/model.py create mode 100644 vllm/models/deepseek_v32/nvidia/mtp.py diff --git a/vllm/models/deepseek_v32/__init__.py b/vllm/models/deepseek_v32/__init__.py new file mode 100644 index 0000000000..1b0aa64262 --- /dev/null +++ b/vllm/models/deepseek_v32/__init__.py @@ -0,0 +1,22 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +"""DeepSeek V3.2 (``deepseek_v32``) model — hardware-isolated entry point. + +DeepSeek V3.2 introduced the DeepSeek Sparse Attention (DSA) architecture: +MLA + a "lightning indexer" that selects the top-k tokens for a sparse MLA +attend. The same model code serves any DSA checkpoint, including GLM-5.2 +(``glm_moe_dsa``), which reuses this architecture. +""" + +from vllm.platforms import current_platform + +if current_platform.is_rocm() or current_platform.is_xpu(): + raise NotImplementedError("deepseek_v32 currently supports NVIDIA SM100 only.") + +from .nvidia.model import DeepseekV32ForCausalLM +from .nvidia.mtp import DeepseekV32MTP + +__all__ = [ + "DeepseekV32ForCausalLM", + "DeepseekV32MTP", +] diff --git a/vllm/models/deepseek_v32/nvidia/__init__.py b/vllm/models/deepseek_v32/nvidia/__init__.py new file mode 100644 index 0000000000..208f01a7cb --- /dev/null +++ b/vllm/models/deepseek_v32/nvidia/__init__.py @@ -0,0 +1,2 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project diff --git a/vllm/models/deepseek_v32/nvidia/attention.py b/vllm/models/deepseek_v32/nvidia/attention.py new file mode 100644 index 0000000000..21b0c2c441 --- /dev/null +++ b/vllm/models/deepseek_v32/nvidia/attention.py @@ -0,0 +1,423 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +from typing import TYPE_CHECKING + +import torch +import torch.nn as nn +from transformers import DeepseekV2Config, DeepseekV3Config + +from vllm.compilation.breakable_cudagraph import eager_break_during_capture +from vllm.config import CacheConfig, VllmConfig +from vllm.distributed import get_tensor_model_parallel_world_size +from vllm.forward_context import get_forward_context +from vllm.model_executor.layers.attention import MLAAttention +from vllm.model_executor.layers.layernorm import LayerNorm, RMSNorm +from vllm.model_executor.layers.linear import ( + ColumnParallelLinear, + MergedColumnParallelLinear, + ReplicatedLinear, + RowParallelLinear, +) +from vllm.model_executor.layers.quantization import QuantizationConfig +from vllm.model_executor.layers.quantization.utils.fp8_utils import ( + per_token_group_quant_fp8, +) +from vllm.model_executor.layers.rotary_embedding import get_rope +from vllm.model_executor.layers.sparse_attn_indexer import SparseAttnIndexer +from vllm.model_executor.models.deepseek_v2 import ( + DeepSeekV2FusedQkvAProjLinear, + DeepseekV32IndexerCache, + yarn_get_mscale, +) +from vllm.model_executor.models.utils import extract_layer_index +from vllm.utils.torch_utils import is_quantized_kv_cache + +if TYPE_CHECKING: + from vllm.model_executor.layers.attention.mla_attention import ( + MLACommonMetadata, + ) + + +class DeepseekV32Indexer(nn.Module): + def __init__( + self, + vllm_config: VllmConfig, + config: DeepseekV2Config | DeepseekV3Config, + hidden_size: int, + q_lora_rank: int, + quant_config: QuantizationConfig | None, + cache_config: CacheConfig | None, + topk_indices_buffer: torch.Tensor | None, + prefix: str = "", + ): + super().__init__() + self.topk_tokens = config.index_topk + self.n_head = config.index_n_heads + self.head_dim = config.index_head_dim + self.rope_dim = config.qk_rope_head_dim + self.q_lora_rank = q_lora_rank + + # No tensor parallel, just replicated. + self.wq_b = ReplicatedLinear( + self.q_lora_rank, + self.head_dim * self.n_head, + bias=False, + quant_config=quant_config, + prefix=f"{prefix}.wq_b", + ) + # Fused wk + weights_proj: single GEMM producing [head_dim + n_head]. + # FP8 wk weights are upcasted to BF16 during loading to keep this fused. + self.wk_weights_proj = MergedColumnParallelLinear( + hidden_size, + [self.head_dim, self.n_head], + bias=False, + quant_config=None, + disable_tp=True, + prefix=f"{prefix}.wk_weights_proj", + ) + self.k_norm = LayerNorm(self.head_dim, eps=1e-6) + self.softmax_scale = self.head_dim**-0.5 + + self.scale_fmt = "ue8m0" + self.quant_block_size = 128 + self.topk_indices_buffer = topk_indices_buffer + + # fp8 naive cache: value in fp8 + fp32 scale per quant_block_size element. + assert cache_config is not None, "DeepSeek V3.2 indexer requires cache_config" + self.k_cache = DeepseekV32IndexerCache( + head_dim=self.head_dim + self.head_dim // self.quant_block_size * 4, + dtype=torch.uint8, + prefix=f"{prefix}.k_cache", + cache_config=cache_config, + ) + self.max_model_len = vllm_config.model_config.max_model_len + self.prefix = prefix + + from vllm.v1.attention.backends.mla.indexer import ( + get_max_prefill_buffer_size, + ) + + self.max_total_seq_len = get_max_prefill_buffer_size(vllm_config) + self.indexer_op = SparseAttnIndexer( + self.k_cache, + self.quant_block_size, + self.scale_fmt, + self.topk_tokens, + self.head_dim, + self.max_model_len, + self.max_total_seq_len, + self.topk_indices_buffer, + ) + + def forward( + self, + hidden_states: torch.Tensor, + qr: torch.Tensor, + positions: torch.Tensor, + rotary_emb: nn.Module, + ) -> torch.Tensor: + q, _ = self.wq_b(qr) + q = q.view(-1, self.n_head, self.head_dim) + + q_pe, q_nope = torch.split( + q, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1 + ) + # Fused wk + weights_proj: one GEMM, then split. + kw, _ = self.wk_weights_proj(hidden_states) + k = kw[:, : self.head_dim] + weights = kw[:, self.head_dim :] + + k = self.k_norm(k) + k_pe, k_nope = torch.split( + k, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1 + ) + + q_pe, k_pe = rotary_emb(positions, q_pe, k_pe.unsqueeze(1)) + # RoPE (NeoX) can introduce extra leading dims; reshape back to flat. + q_pe = q_pe.reshape(-1, self.n_head, self.rope_dim) + k_pe = k_pe.reshape(-1, 1, self.rope_dim) + + q = torch.cat([q_pe, q_nope], dim=-1) + k = torch.cat([k_pe.squeeze(-2), k_nope], dim=-1) + + # Only quant q here; k quant is fused with cache insertion. + q = q.view(-1, self.head_dim) + q_fp8, q_scale = per_token_group_quant_fp8( + q, + self.quant_block_size, + column_major_scales=False, + use_ue8m0=self.scale_fmt is not None, + ) + q_fp8 = q_fp8.view(-1, self.n_head, self.head_dim) + q_scale = q_scale.view(-1, self.n_head, 1) + + weights = ( + weights.unsqueeze(-1) * q_scale * self.softmax_scale * self.n_head**-0.5 + ) + weights = weights.squeeze(-1) + + return self.indexer_op(hidden_states, q_fp8, k, weights) + + +class DeepseekV32Attention(MLAAttention): + def __init__( + self, + vllm_config: VllmConfig, + config: DeepseekV2Config | DeepseekV3Config, + prefix: str, + topk_indices_buffer: torch.Tensor | None = None, + ) -> None: + quant_config = vllm_config.quant_config + cache_config = vllm_config.cache_config + + hidden_size = config.hidden_size + qk_nope_head_dim = config.qk_nope_head_dim + qk_rope_head_dim = config.qk_rope_head_dim + v_head_dim = config.v_head_dim + q_lora_rank = config.q_lora_rank + kv_lora_rank = config.kv_lora_rank + num_heads = config.num_attention_heads + + tp_size = get_tensor_model_parallel_world_size() + assert num_heads % tp_size == 0 + num_local_heads = num_heads // tp_size + qk_head_dim = qk_nope_head_dim + qk_rope_head_dim + scaling = qk_head_dim**-0.5 + max_position_embeddings = getattr(config, "max_position_embeddings", 8192) + + # DSA checkpoints may use plain ("default") or yarn-scaled RoPE. + if config.rope_parameters["rope_type"] != "default": + config.rope_parameters["rope_type"] = ( + "deepseek_yarn" + if config.rope_parameters.get("apply_yarn_scaling", True) + else "deepseek_llama_scaling" + ) + if config.rope_parameters["rope_type"] == "deepseek_yarn": + mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False) + scaling_factor = config.rope_parameters["factor"] + mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim)) + scaling = scaling * mscale * mscale + + # DSA "shared indexer" pattern: only some layers carry an indexer; the + # rest reuse the top-k written by the previous indexer layer into the + # shared topk_indices_buffer. DeepSeek-V3.2 builds it on every layer + # (index_topk_freq defaults to 1); GLM-5.2 uses index_topk_freq=4 so + # only layers [0,1,2,6,10,...] (+ MTP) carry one. + layer_id = extract_layer_index(prefix) + index_topk_freq = getattr(config, "index_topk_freq", 1) + index_topk_pattern = getattr(config, "index_topk_pattern", None) + index_skip_topk_offset = getattr(config, "index_skip_topk_offset", 2) + if index_topk_pattern is None: + skip_topk = ( + max(layer_id - index_skip_topk_offset + 1, 0) % index_topk_freq != 0 + ) + elif 0 <= layer_id < len(index_topk_pattern): + skip_topk = index_topk_pattern[layer_id] == "S" + else: + skip_topk = False + # MTP/nextn layers always build a full indexer (they toggle at runtime). + num_hidden_layers = getattr(config, "num_hidden_layers", None) + is_mtp_layer = num_hidden_layers is not None and layer_id >= num_hidden_layers + + # Build kv_b_proj + indexer first; they are passed to MLAAttention.__init__ + # (which runs nn.Module.__init__ and registers them). + kv_b_proj = ColumnParallelLinear( + kv_lora_rank, + num_heads * (qk_nope_head_dim + v_head_dim), + bias=False, + quant_config=quant_config, + prefix=f"{prefix}.kv_b_proj", + ) + indexer = None + if not skip_topk or is_mtp_layer: + indexer = DeepseekV32Indexer( + vllm_config, + config, + hidden_size, + q_lora_rank, + quant_config, + cache_config, + topk_indices_buffer, + prefix=f"{prefix}.indexer", + ) + + # Set up the MLA engine (impl, KV cache, scales, backend, registration, + # and process_weights_after_loading) via the MLAAttention base. + super().__init__( + num_heads=num_local_heads, + scale=scaling, + qk_nope_head_dim=qk_nope_head_dim, + qk_rope_head_dim=qk_rope_head_dim, + v_head_dim=v_head_dim, + q_lora_rank=q_lora_rank, + kv_lora_rank=kv_lora_rank, + kv_b_proj=kv_b_proj, + cache_config=cache_config, + quant_config=quant_config, + prefix=f"{prefix}.attn", + use_sparse=True, + indexer=indexer, + topk_indices_buffer=topk_indices_buffer, + ) + + self.num_local_heads = num_local_heads + self.qk_head_dim = qk_head_dim + self.indexer = indexer + # Runtime toggle for index_share_for_mtp_iteration: MTP draft step 0 + # computes the top-k, steps 1+ set this True to reuse it. + self.skip_topk = False + # Whether the paged KV cache must be viewed as fp8 before the attention + # (per-tensor fp8; the fp8_ds_mla layout is read as uint8). + self._fp8_kv_needs_view = ( + is_quantized_kv_cache(self.kv_cache_dtype) + and self.kv_cache_dtype != "fp8_ds_mla" + ) + # Whether the backend takes an fp8-quantized query (FlashInfer sparse) + # vs the (ql_nope, q_pe) tuple (FlashMLA sparse). + self._use_concat_quant = ( + is_quantized_kv_cache(self.kv_cache_dtype) + and self.impl.supports_quant_query_input + ) + + # Remaining MLA projections (registered on this module). + self.fused_qkv_a_proj = DeepSeekV2FusedQkvAProjLinear( + hidden_size, + [q_lora_rank, kv_lora_rank + qk_rope_head_dim], + quant_config=quant_config, + prefix=f"{prefix}.fused_qkv_a_proj", + ) + self.q_a_layernorm = RMSNorm(q_lora_rank, eps=config.rms_norm_eps) + self.q_b_proj = ColumnParallelLinear( + q_lora_rank, + num_heads * qk_head_dim, + bias=False, + quant_config=quant_config, + prefix=f"{prefix}.q_b_proj", + ) + self.kv_a_layernorm = RMSNorm(kv_lora_rank, eps=config.rms_norm_eps) + self.o_proj = RowParallelLinear( + num_heads * v_head_dim, + hidden_size, + bias=False, + quant_config=quant_config, + prefix=f"{prefix}.o_proj", + ) + + self.rotary_emb = get_rope( + qk_rope_head_dim, + max_position=max_position_embeddings, + rope_parameters=config.rope_parameters, + is_neox_style=False, + ) + # Lightning indexer uses its own RoPE; interleave maps to non-NeoX. + self.indexer_rope_emb = get_rope( + qk_rope_head_dim, + max_position=max_position_embeddings, + rope_parameters=config.rope_parameters, + is_neox_style=not getattr(config, "indexer_rope_interleave", False), + ) + + def forward( # type: ignore[override] + self, + positions: torch.Tensor, + hidden_states: torch.Tensor, + ) -> torch.Tensor: + qkv_lora = self.fused_qkv_a_proj(hidden_states)[0] + q_c, kv_lora = qkv_lora.split( + [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], dim=-1 + ) + q_c = self.q_a_layernorm(q_c) + q = self.q_b_proj(q_c)[0] + + kv_c, k_pe = kv_lora.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1) + kv_c_normed = self.kv_a_layernorm(kv_c) + + q = q.view(-1, self.num_local_heads, self.qk_head_dim) + k_pe = k_pe.unsqueeze(1) + q[..., self.qk_nope_head_dim :], k_pe = self.rotary_emb( + positions, q[..., self.qk_nope_head_dim :], k_pe + ) + + num_tokens = hidden_states.shape[0] + q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1) + q_nope = q_nope.transpose(0, 1) # (N, B, P) + ql_nope = torch.bmm(q_nope, self.W_UK_T).transpose(0, 1) # (B, N, L) + + # Lightning indexer writes the top-k indices into the shared buffer. + # "Shared" layers (indexer is None) reuse the top-k from the previous + # indexer layer already sitting in the buffer. + if self.indexer is not None and not self.skip_topk: + self.indexer(hidden_states, q_c, positions, self.indexer_rope_emb) # type: ignore[operator] + + attn_latent = torch.empty( + (num_tokens, self.num_local_heads, self.kv_lora_rank), + dtype=q.dtype, + device=q.device, + ) + self._sparse_attention(kv_c_normed, k_pe, ql_nope, q_pe, attn_latent) + + # V up-projection + output projection are metadata-independent GEMMs and + # stay captured. + output = torch.empty( + (num_tokens, self.num_local_heads * self.v_head_dim), + dtype=q.dtype, + device=q.device, + ) + self._v_up_proj(attn_latent, out=output) + return self.o_proj(output)[0] + + @eager_break_during_capture + def _sparse_attention( + self, + kv_c_normed: torch.Tensor, + k_pe: torch.Tensor, + ql_nope: torch.Tensor, + q_pe: torch.Tensor, + attn_latent: torch.Tensor, + ) -> None: + forward_context = get_forward_context() + attn_metadata_raw = forward_context.attn_metadata + attn_metadata: MLACommonMetadata | None + if isinstance(attn_metadata_raw, dict): + attn_metadata = attn_metadata_raw[self.layer_name] # type: ignore[assignment] + elif isinstance(attn_metadata_raw, list): + # Speculative decoding: [0] is the base-model metadata dict. + attn_metadata = attn_metadata_raw[0][self.layer_name] # type: ignore[assignment] + else: + attn_metadata = attn_metadata_raw + + slot_mapping = forward_context.slot_mapping + assert isinstance(slot_mapping, dict) + self.impl.do_kv_cache_update( # type: ignore[attr-defined] + kv_c_normed, + k_pe, + self.kv_cache, + slot_mapping.get(self.layer_name), + self.kv_cache_dtype, + self._k_scale, + ) + + if attn_metadata is None: + # Profile / warmup: zero-fill for DP+EP determinism. + attn_latent.zero_() + return + + num_actual = attn_metadata.num_actual_tokens + kv_cache = self.kv_cache + if self._fp8_kv_needs_view: + kv_cache = kv_cache.view(torch.float8_e4m3fn) + + ql_nope = ql_nope[:num_actual] + q_pe = q_pe[:num_actual] + # FlashInfer sparse takes a single fp8-quantized query; FlashMLA sparse + # takes the (ql_nope, q_pe) tuple and concatenates internally. + mqa_q: torch.Tensor | tuple[torch.Tensor, torch.Tensor] + if self._use_concat_quant: + mqa_q = self._decode_concat_quant_fp8_op(ql_nope, q_pe, self._q_scale) + else: + mqa_q = (ql_nope, q_pe) + + attn_out, _ = self.impl.forward_mqa(mqa_q, kv_cache, attn_metadata, self) # type: ignore[attr-defined] + attn_latent[:num_actual] = attn_out.view( + num_actual, self.num_local_heads, self.kv_lora_rank + ) diff --git a/vllm/models/deepseek_v32/nvidia/model.py b/vllm/models/deepseek_v32/nvidia/model.py new file mode 100644 index 0000000000..dd9e1d65ea --- /dev/null +++ b/vllm/models/deepseek_v32/nvidia/model.py @@ -0,0 +1,333 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +import typing +from collections.abc import Callable, Iterable +from itertools import islice + +import torch + +from vllm.config import VllmConfig +from vllm.distributed import get_pp_group +from vllm.model_executor.layers.fused_moe import ( + fused_moe_make_expert_params_mapping, +) +from vllm.model_executor.layers.layernorm import RMSNorm +from vllm.model_executor.layers.vocab_parallel_embedding import ( + VocabParallelEmbedding, +) +from vllm.model_executor.model_loader.weight_utils import ( + default_weight_loader, + maybe_remap_kv_scale_name, +) +from vllm.model_executor.models.deepseek_v2 import ( + DeepseekV2ForCausalLM, + DeepseekV2MLP, + DeepseekV2MoE, + _try_load_fp8_indexer_wk, + get_spec_layer_idx_from_weight_name, +) +from vllm.model_executor.models.utils import ( + PPMissingLayer, + get_pp_missing_layer_names, + is_pp_missing_parameter, + make_empty_intermediate_tensors_factory, + make_layers, +) +from vllm.sequence import IntermediateTensors + +from .attention import DeepseekV32Attention + + +class DeepseekV32DecoderLayer(torch.nn.Module): + def __init__( + self, + vllm_config: VllmConfig, + prefix: str, + config=None, + topk_indices_buffer: torch.Tensor | None = None, + ) -> None: + super().__init__() + + if config is None: + config = vllm_config.model_config.hf_config + quant_config = vllm_config.quant_config + parallel_config = vllm_config.parallel_config + + self.hidden_size = config.hidden_size + moe_layer_freq = getattr(config, "moe_layer_freq", 1) + layer_idx = int(prefix.split(sep=".")[-1]) + self.layer_idx = layer_idx + self.use_mha = False + + self.self_attn = DeepseekV32Attention( + vllm_config=vllm_config, + config=config, + prefix=f"{prefix}.self_attn", + topk_indices_buffer=topk_indices_buffer, + ) + + if ( + config.n_routed_experts is not None + and layer_idx >= config.first_k_dense_replace + and layer_idx % moe_layer_freq == 0 + ): + self.mlp = DeepseekV2MoE( + config=config, + parallel_config=parallel_config, + quant_config=quant_config, + prefix=f"{prefix}.mlp", + ) + else: + self.mlp = DeepseekV2MLP( + 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 + ) + self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0) + + def forward( + self, + positions: torch.Tensor, + hidden_states: torch.Tensor, + residual: torch.Tensor | None, + ) -> tuple[torch.Tensor, 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) + hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) + hidden_states = self.mlp(hidden_states) + return hidden_states, residual + + +class DeepseekV32Model(torch.nn.Module): + fall_back_to_pt_during_load = False + + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): + super().__init__() + + config = vllm_config.model_config.hf_config + quant_config = vllm_config.quant_config + self.config = config + from vllm.platforms import current_platform + + self.device = current_platform.device_type + + self.vocab_size = config.vocab_size + # DSA is always sparse (has index_topk); allocate the shared top-k + # buffer the indexer writes and the sparse MLA backend reads. + self.is_v32 = True + topk_indices_buffer = torch.empty( + vllm_config.scheduler_config.max_num_batched_tokens, + config.index_topk, + dtype=torch.int32, + device=self.device, + ) + + if get_pp_group().is_first_rank: + self.embed_tokens = VocabParallelEmbedding( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=f"{prefix}.embed_tokens", + ) + else: + self.embed_tokens = PPMissingLayer() + + self.start_layer, self.end_layer, self.layers = make_layers( + config.num_hidden_layers, + lambda prefix: DeepseekV32DecoderLayer( + vllm_config=vllm_config, + prefix=prefix, + topk_indices_buffer=topk_indices_buffer, + ), + prefix=f"{prefix}.layers", + ) + + if get_pp_group().is_last_rank: + self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + else: + self.norm = PPMissingLayer() + self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( + ["hidden_states", "residual"], config.hidden_size + ) + + self.aux_hidden_state_layers = tuple[int, ...]() + self.num_redundant_experts = ( + vllm_config.parallel_config.eplb_config.num_redundant_experts + ) + + def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: + return self.embed_tokens(input_ids) + + def forward( + self, + input_ids: torch.Tensor | None, + positions: torch.Tensor, + intermediate_tensors: IntermediateTensors | None = None, + inputs_embeds: torch.Tensor | None = None, + ) -> torch.Tensor | IntermediateTensors: + if get_pp_group().is_first_rank: + if inputs_embeds is not None: + hidden_states = inputs_embeds + else: + assert input_ids is not None + hidden_states = self.embed_input_ids(input_ids) + residual = None + else: + assert intermediate_tensors is not None + hidden_states = intermediate_tensors["hidden_states"] + residual = intermediate_tensors["residual"] + + aux_hidden_states = [] + for idx, layer in enumerate( + islice(self.layers, self.start_layer, self.end_layer), + start=self.start_layer, + ): + if idx in self.aux_hidden_state_layers: + aux_hidden_states.append(hidden_states + residual) + hidden_states, residual = layer(positions, hidden_states, residual) + + if not get_pp_group().is_last_rank: + return IntermediateTensors( + {"hidden_states": hidden_states, "residual": residual} + ) + + 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]: + # DSA-only: MLA (fused_qkv_a_proj) + the fused indexer wk/weights_proj + + # routed experts. No MHA (qkv_proj) or ROCm shared-expert-fusion paths. + stacked_params_mapping = [ + ("gate_up_proj", "gate_proj", 0), + ("gate_up_proj", "up_proj", 1), + ("fused_qkv_a_proj", "q_a_proj", 0), + ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1), + ("wk_weights_proj", "wk", 0), + ("wk_weights_proj", "weights_proj", 1), + ] + expert_params_mapping = fused_moe_make_expert_params_mapping( + self, + ckpt_gate_proj_name="gate_proj", + ckpt_down_proj_name="down_proj", + ckpt_up_proj_name="up_proj", + num_experts=self.config.n_routed_experts, + num_redundant_experts=self.num_redundant_experts, + ) + + pp_missing_layer_names = get_pp_missing_layer_names(self) + params_dict = dict(self.named_parameters()) + loaded_params: set[str] = set() + _pending_wk_fp8: dict = {} + for name, loaded_weight in weights: + if "rotary_emb.inv_freq" in name: + continue + # MTP / nextn layers are loaded by the MTP model, not here. + if get_spec_layer_idx_from_weight_name(self.config, name) is not None: + continue + if _try_load_fp8_indexer_wk( + name, + loaded_weight, + _pending_wk_fp8, + params_dict, + loaded_params, + pp_missing_layer_names, + ): + continue + + for param_name, weight_name, shard_id in stacked_params_mapping: + if weight_name not in name: + continue + # Experts are handled below; skip here before the name rewrite. + if ("mlp.experts." in name) and name not in params_dict: + continue + name_mapped = name.replace(weight_name, param_name) + if ( + param_name == "fused_qkv_a_proj" + ) and name_mapped not in params_dict: + continue + name = name_mapped + if name.endswith(".bias") and name not in params_dict: + continue + if is_pp_missing_parameter(name, self): + continue + param = params_dict[name] + param.weight_loader(param, loaded_weight, shard_id) + break + else: + is_expert_weight = False + for mapping in expert_params_mapping: + param_name, weight_name, expert_id, shard_id = mapping # type: ignore[assignment] + if weight_name not in name: + continue + is_expert_weight = True + name_mapped = name.replace(weight_name, param_name) + if is_pp_missing_parameter(name_mapped, self): + continue + param = params_dict[name_mapped] + weight_loader = typing.cast( + Callable[..., bool], param.weight_loader + ) + success = weight_loader( + param, + loaded_weight, + name_mapped, + shard_id=shard_id, + expert_id=expert_id, + return_success=True, + ) + if success: + name = name_mapped + break + else: + if is_expert_weight: + continue + if name.endswith(".bias") and name not in params_dict: + continue + name = maybe_remap_kv_scale_name(name, params_dict) # type: ignore[assignment] + if name is None: + continue + if is_pp_missing_parameter(name, self): + continue + param = params_dict[name] + loader = getattr(param, "weight_loader", default_weight_loader) + loader(param, loaded_weight) + loaded_params.add(name) + return loaded_params + + +class DeepseekV32ForCausalLM(DeepseekV2ForCausalLM): + """DSA causal LM — DeepSeek V2/V3 orchestration with the DSA backbone. + + Serves DeepSeek V3.2 and any architecture reusing DSA (e.g. GLM-5.2). + """ + + model_cls = DeepseekV32Model + + def set_moe_parameters(self): + # Same as the base, but keyed on the MoE block type rather than the + # decoder-layer type (DeepseekV32DecoderLayer is a plain nn.Module). + self.expert_weights = [] + self.num_expert_groups = getattr(self.config, "n_group", 1) + 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, DeepseekV2MoE): + example_moe = layer.mlp + self.moe_mlp_layers.append(layer.mlp) + self.moe_layers.append(layer.mlp.experts) + self.extract_moe_parameters(example_moe) diff --git a/vllm/models/deepseek_v32/nvidia/mtp.py b/vllm/models/deepseek_v32/nvidia/mtp.py new file mode 100644 index 0000000000..482ebecc52 --- /dev/null +++ b/vllm/models/deepseek_v32/nvidia/mtp.py @@ -0,0 +1,390 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import typing +from collections.abc import Callable, Iterable + +import torch +import torch.nn as nn + +from vllm._aiter_ops import rocm_aiter_ops +from vllm.config import VllmConfig +from vllm.model_executor.layers.fused_moe import ( + fused_moe_make_expert_params_mapping, +) +from vllm.model_executor.layers.layernorm import RMSNorm +from vllm.model_executor.layers.logits_processor import LogitsProcessor +from vllm.model_executor.layers.vocab_parallel_embedding import ( + VocabParallelEmbedding, +) +from vllm.model_executor.model_loader.weight_utils import ( + default_weight_loader, + maybe_remap_kv_scale_name, +) +from vllm.model_executor.models.deepseek_mtp import SharedHead +from vllm.model_executor.models.deepseek_v2 import ( + DeepseekV2MixtureOfExperts, + DeepseekV2MoE, + _try_load_fp8_indexer_wk, + get_spec_layer_idx_from_weight_name, +) +from vllm.model_executor.models.utils import ( + get_pp_missing_layer_names, + maybe_prefix, +) +from vllm.platforms import current_platform +from vllm.sequence import IntermediateTensors + +from .model import DeepseekV32DecoderLayer + + +class DeepseekV32MultiTokenPredictorLayer(nn.Module): + def __init__(self, vllm_config: VllmConfig, prefix: str) -> None: + super().__init__() + assert vllm_config.speculative_config is not None + config = vllm_config.speculative_config.draft_model_config.hf_config + self.config = config + quant_config = vllm_config.quant_config + + self.enorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.hnorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.eh_proj = nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False) + + topk_indices_buffer = torch.empty( + vllm_config.scheduler_config.max_num_batched_tokens, + config.index_topk, + dtype=torch.int32, + device=current_platform.device_type, + ) + self.shared_head = SharedHead( + config=config, prefix=prefix, quant_config=quant_config + ) + self.mtp_block = DeepseekV32DecoderLayer( + vllm_config, + prefix, + config=config, + topk_indices_buffer=topk_indices_buffer, + ) + + def forward( + self, + input_ids: torch.Tensor, + positions: torch.Tensor, + previous_hidden_states: torch.Tensor, + inputs_embeds: torch.Tensor | None = None, + spec_step_index: int = 0, + ) -> torch.Tensor: + assert inputs_embeds is not None + inputs_embeds = torch.where(positions.unsqueeze(-1) == 0, 0, inputs_embeds) + inputs_embeds = self.enorm(inputs_embeds) + previous_hidden_states = self.hnorm(previous_hidden_states) + hidden_states = self.eh_proj( + torch.cat([inputs_embeds, previous_hidden_states], dim=-1) + ) + hidden_states, residual = self.mtp_block( + positions=positions, hidden_states=hidden_states, residual=None + ) + # Return the pre-final-norm recycle hidden (re-fed as the next spec + # step's previous_hidden_states); shared_head norm is applied in + # compute_logits. Matches the V2-runner / deepseek_v4 MTP contract. + return residual + hidden_states + + +class DeepseekV32MultiTokenPredictor(nn.Module): + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): + super().__init__() + config = vllm_config.model_config.hf_config + self.mtp_start_layer_idx = config.num_hidden_layers + self.num_mtp_layers = config.num_nextn_predict_layers + self.layers = torch.nn.ModuleDict( + { + str(idx): DeepseekV32MultiTokenPredictorLayer( + vllm_config, f"{prefix}.layers.{idx}" + ) + for idx in range( + self.mtp_start_layer_idx, + self.mtp_start_layer_idx + self.num_mtp_layers, + ) + } + ) + self.embed_tokens = VocabParallelEmbedding( + config.vocab_size, + config.hidden_size, + prefix=maybe_prefix(prefix, "embed_tokens"), + ) + self.logits_processor = LogitsProcessor(config.vocab_size) + + def set_skip_topk(self, skip: bool): + # index_share_for_mtp_iteration: step 0 computes top-k, steps 1+ reuse. + for layer in self.layers.values(): + self_attn = getattr(layer.mtp_block, "self_attn", None) + if self_attn is not None and hasattr(self_attn, "skip_topk"): + self_attn.skip_topk = skip + + 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, + previous_hidden_states: torch.Tensor, + inputs_embeds: torch.Tensor | None = None, + spec_step_idx: int = 0, + ) -> torch.Tensor: + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + current_step_idx = spec_step_idx % self.num_mtp_layers + return self.layers[str(self.mtp_start_layer_idx + current_step_idx)]( + input_ids, + positions, + previous_hidden_states, + inputs_embeds, + current_step_idx, + ) + + def compute_logits( + self, + hidden_states: torch.Tensor, + spec_step_idx: int = 0, + ) -> torch.Tensor: + current_step_idx = spec_step_idx % self.num_mtp_layers + mtp_layer = self.layers[str(self.mtp_start_layer_idx + current_step_idx)] + return self.logits_processor( + mtp_layer.shared_head.head, mtp_layer.shared_head(hidden_states) + ) + + +class DeepseekV32MTP(nn.Module, DeepseekV2MixtureOfExperts): + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): + super().__init__() + self.config = vllm_config.model_config.hf_config + self.quant_config = vllm_config.quant_config + self.model = DeepseekV32MultiTokenPredictor( + vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") + ) + self.set_moe_parameters() + + def set_moe_parameters(self): + self.expert_weights = [] + self.num_moe_layers = self.config.num_nextn_predict_layers + self.num_expert_groups = self.config.n_group + self.moe_layers = [] + self.moe_mlp_layers = [] + example_moe = None + for layer in self.model.layers.values(): + mlp = layer.mtp_block.mlp + if isinstance(mlp, DeepseekV2MoE): + example_moe = mlp + self.moe_mlp_layers.append(mlp) + self.moe_layers.append(mlp.experts) + 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 | None, + positions: torch.Tensor, + hidden_states: torch.Tensor, + intermediate_tensors: IntermediateTensors | None = None, + inputs_embeds: torch.Tensor | None = None, + spec_step_idx: int = 0, + ) -> torch.Tensor: + return self.model( + input_ids, positions, hidden_states, inputs_embeds, spec_step_idx + ) + + def compute_logits( + self, + hidden_states: torch.Tensor, + spec_step_idx: int = 0, + ) -> torch.Tensor | None: + return self.model.compute_logits(hidden_states, spec_step_idx) + + def _rewrite_spec_layer_name(self, spec_layer: int, name: str) -> str: + spec_layer_weight_names = [ + "embed_tokens", + "enorm", + "hnorm", + "eh_proj", + "shared_head", + ] + shared_weight_names = ["embed_tokens"] + spec_layer_weight = False + shared_weight = False + for weight_name in spec_layer_weight_names: + if weight_name in name: + spec_layer_weight = True + if weight_name in shared_weight_names: + shared_weight = True + break + if not spec_layer_weight: + name = name.replace( + f"model.layers.{spec_layer}.", f"model.layers.{spec_layer}.mtp_block." + ) + elif shared_weight: + name = name.replace(f"model.layers.{spec_layer}.", "model.") + return name + + def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: + rocm_aiter_moe_shared_expert_enabled = ( + rocm_aiter_ops.is_fusion_moe_shared_experts_enabled() + ) + stacked_params_mapping = [ + ("gate_up_proj", "gate_proj", 0), + ("gate_up_proj", "up_proj", 1), + ("fused_qkv_a_proj", "q_a_proj", 0), + ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1), + ("wk_weights_proj", "wk", 0), + ("wk_weights_proj", "weights_proj", 1), + ] + expert_params_mapping = fused_moe_make_expert_params_mapping( + self, + ckpt_gate_proj_name="gate_proj", + ckpt_down_proj_name="down_proj", + ckpt_up_proj_name="up_proj", + num_experts=self.config.n_routed_experts + + ( + self.config.n_shared_experts + if rocm_aiter_moe_shared_expert_enabled + else 0 + ), + ) + + pp_missing_layer_names = get_pp_missing_layer_names(self) + params_dict = dict(self.named_parameters()) + loaded_params: set[str] = set() + _pending_wk_fp8: dict = {} + for name, loaded_weight in weights: + if "rotary_emb.inv_freq" in name: + continue + spec_layer = get_spec_layer_idx_from_weight_name(self.config, name) + if spec_layer is None: + continue + is_fusion_moe_shared_experts_layer = ( + rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name) + ) + name = self._rewrite_spec_layer_name(spec_layer, name) + + if _try_load_fp8_indexer_wk( + name, + loaded_weight, + _pending_wk_fp8, + params_dict, + loaded_params, + pp_missing_layer_names, + ): + continue + + for param_name, weight_name, shard_id in stacked_params_mapping: + if weight_name not in name: + continue + if ("mlp.experts." in name) and name not in params_dict: + continue + if is_fusion_moe_shared_experts_layer: + continue + name_mapped = name.replace(weight_name, param_name) + if ( + param_name == "fused_qkv_a_proj" + ) and name_mapped not in params_dict: + continue + else: + name = name_mapped + if name.endswith(".bias") and name not in params_dict: + continue + param = params_dict[name] + weight_loader = param.weight_loader + weight_loader(param, loaded_weight, shard_id) + break + else: + num_chunks = 1 + if is_fusion_moe_shared_experts_layer: + num_chunks = getattr(self.config, "n_shared_experts", 1) or 1 + split_dim = ( + 1 + if ("down_proj.weight" in name and loaded_weight.ndim > 1) + else 0 + ) + total = loaded_weight.shape[split_dim] + assert total % num_chunks == 0 + chunk_size = total // num_chunks + + for j in range(num_chunks): + chunk_name = name + weight_to_load = loaded_weight + if is_fusion_moe_shared_experts_layer: + chunk_slice = slice(j * chunk_size, (j + 1) * chunk_size) + if loaded_weight.ndim == 1: + weight_to_load = loaded_weight[chunk_slice] + elif split_dim == 0: + weight_to_load = loaded_weight[chunk_slice, :] + else: + weight_to_load = loaded_weight[:, chunk_slice] + chunk_name = name.replace( + "mlp.shared_experts", + f"mlp.experts.{self.config.n_routed_experts + j}", + ) + + is_expert_weight = False + for mapping in expert_params_mapping: + param_name, weight_name, expert_id, shard_id = mapping # type: ignore[assignment] + if weight_name not in chunk_name: + continue + is_expert_weight = True + name_mapped = chunk_name.replace(weight_name, param_name) + param = params_dict[name_mapped] + weight_loader = typing.cast( + Callable[..., bool], param.weight_loader + ) + success = weight_loader( + param, + weight_to_load, + name_mapped, + shard_id=shard_id, + expert_id=expert_id, + return_success=True, + ) + if success: + if not is_fusion_moe_shared_experts_layer: + name = name_mapped + else: + loaded_params.add(name_mapped) + break + else: + if is_expert_weight: + continue + if name.endswith(".bias") and name not in params_dict: + continue + name = maybe_remap_kv_scale_name(name, params_dict) # type: ignore[assignment] + if name is None: + continue + if ( + spec_layer != self.model.mtp_start_layer_idx + and ".layers" not in name + ): + continue + param = params_dict[name] + weight_loader = getattr( + param, "weight_loader", default_weight_loader + ) + weight_loader(param, loaded_weight) + if not is_fusion_moe_shared_experts_layer: + loaded_params.add(name) + + loaded_layers: set[int] = set() + for param_name in loaded_params: + spec_layer = get_spec_layer_idx_from_weight_name(self.config, param_name) + if spec_layer is not None: + loaded_layers.add(spec_layer) + for layer_idx in range( + self.model.mtp_start_layer_idx, + self.model.mtp_start_layer_idx + self.model.num_mtp_layers, + ): + if layer_idx not in loaded_layers: + raise ValueError( + f"MTP speculative decoding layer {layer_idx} weights " + f"missing from checkpoint." + ) + return loaded_params