Deepseek v3 (#11502)

Signed-off-by: mgoin <michael@neuralmagic.com>
Co-authored-by: mgoin <michael@neuralmagic.com>
Co-authored-by: robertgshaw2-neuralmagic <rshaw@neuralmagic.com>

csrc/moe/moe_align_sum_kernels.cu

@@ -113,6 +113,92 @@ __global__ void moe_align_block_size_kernel(scalar_t* __restrict__ topk_ids,
   }
 }
 
+// TODO(simon): this is temporarily adapted from
+// https://github.com/sgl-project/sglang/commit/31548116a8dc8c6df7e146e0587335a59fc5b9d7
+// we did this to unblock Deepseek V3 but there should be a better
+// implementation to manage shared memory.
+template <typename scalar_t>
+__global__ void moe_align_block_size_global_mem_kernel(
+    scalar_t* __restrict__ topk_ids, int32_t* sorted_token_ids,
+    int32_t* expert_ids, int32_t* total_tokens_post_pad, int32_t num_experts,
+    int32_t block_size, size_t numel, int32_t* tokens_cnts, int32_t* cumsum) {
+  const size_t tokens_per_thread = CEILDIV(numel, blockDim.x);
+  const size_t start_idx = threadIdx.x * tokens_per_thread;
+
+  for (int i = 0; i < num_experts; ++i) {
+    tokens_cnts[index(num_experts, threadIdx.x + 1, i)] = 0;
+  }
+
+  /**
+   * In the first step we compute token_cnts[thread_index + 1][expert_index],
+   * which counts how many tokens in the token shard of thread_index are
+   * assigned to expert expert_index.
+   */
+  for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+    ++tokens_cnts[index(num_experts, threadIdx.x + 1, topk_ids[i])];
+  }
+
+  __syncthreads();
+
+  // For each expert we accumulate the token counts from the different threads.
+  if (threadIdx.x < num_experts) {
+    tokens_cnts[index(num_experts, 0, threadIdx.x)] = 0;
+    for (int i = 1; i <= blockDim.x; ++i) {
+      tokens_cnts[index(num_experts, i, threadIdx.x)] +=
+          tokens_cnts[index(num_experts, i - 1, threadIdx.x)];
+    }
+  }
+
+  __syncthreads();
+
+  // We accumulate the token counts of all experts in thread 0.
+  if (threadIdx.x == 0) {
+    cumsum[0] = 0;
+    for (int i = 1; i <= num_experts; ++i) {
+      cumsum[i] = cumsum[i - 1] +
+                  CEILDIV(tokens_cnts[index(num_experts, blockDim.x, i - 1)],
+                          block_size) *
+                      block_size;
+    }
+    *total_tokens_post_pad = cumsum[num_experts];
+  }
+
+  __syncthreads();
+
+  /**
+   * For each expert, each thread processes the tokens of the corresponding
+   * blocks and stores the corresponding expert_id for each block.
+   */
+  if (threadIdx.x < num_experts) {
+    for (int i = cumsum[threadIdx.x]; i < cumsum[threadIdx.x + 1];
+         i += block_size) {
+      expert_ids[i / block_size] = threadIdx.x;
+    }
+  }
+
+  /**
+   * Each thread processes a token shard, calculating the index of each token
+   * after sorting by expert number. Given the example topk_ids =
+   * [0,1,2,1,2,3,0,3,4] and block_size = 4, then the output would be [0, 6, *,
+   * *, 1, 3, *, *, 2, 4, *, *, 5, 7, *, *, 8, *, *, *], where * represents a
+   * padding value(preset in python).
+   */
+  for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+    int32_t expert_id = topk_ids[i];
+    /** The cumsum[expert_id] stores the starting index of the tokens that the
+     * expert with expert_id needs to process, and
+     * tokens_cnts[threadIdx.x][expert_id] stores the indices of the tokens
+     * processed by the expert with expert_id within the current thread's token
+     * shard.
+     */
+    int32_t rank_post_pad =
+        tokens_cnts[index(num_experts, threadIdx.x, expert_id)] +
+        cumsum[expert_id];
+    sorted_token_ids[rank_post_pad] = i;
+    ++tokens_cnts[index(num_experts, threadIdx.x, expert_id)];
+  }
+}
+
 template <typename scalar_t, int TOPK>
 __global__ void moe_sum_kernel(
     scalar_t* __restrict__ out,          // [..., d]
@@ -137,25 +223,61 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
                           torch::Tensor experts_ids,
                           torch::Tensor num_tokens_post_pad) {
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  VLLM_DISPATCH_INTEGRAL_TYPES(
-      topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
-        // calc needed amount of shared mem for `tokens_cnts` and `cumsum`
-        // tensors
-        const int32_t num_thread = max((int32_t)num_experts, WARP_SIZE);
-        const int32_t shared_mem =
-            ((num_thread + 1) * num_experts + (num_experts + 1)) *
-            sizeof(int32_t);
-
-        // set dynamic shared mem
-        auto kernel = vllm::moe::moe_align_block_size_kernel<scalar_t>;
-        AT_CUDA_CHECK(VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(
-            (void*)kernel, shared_mem));
-        kernel<<<1, num_thread, shared_mem, stream>>>(
-            topk_ids.data_ptr<scalar_t>(), sorted_token_ids.data_ptr<int32_t>(),
-            experts_ids.data_ptr<int32_t>(),
-            num_tokens_post_pad.data_ptr<int32_t>(), num_experts, block_size,
-            topk_ids.numel());
-      });
+
+  // If we have very large number of experts, we can no longer use shared
+  // memory.
+  // TODO(simon): the right solution should be calculating the exact right
+  // amount of shared memory and use that. The num_experts >= 256 is just a
+  // temporary solution to unblock Deepseek V3.
+  if (num_experts >= 256) {
+    VLLM_DISPATCH_INTEGRAL_TYPES(
+        topk_ids.scalar_type(), "moe_align_block_size_global_mem_kernel", [&] {
+          // calc needed amount of shared mem for `tokens_cnts` and `cumsum`
+          // tensors
+          const int32_t num_thread = max((int32_t)num_experts, WARP_SIZE);
+
+          const int32_t mem_tokens_cnts =
+              ((num_experts + 1) * num_experts) * sizeof(int32_t);
+          const int32_t mem_cumsum = (num_experts + 1) * sizeof(int32_t);
+          // allocate global memory
+          int32_t* tokens_cnts;
+          int32_t* cumsum;
+          cudaMalloc(&tokens_cnts, mem_tokens_cnts);
+          cudaMalloc(&cumsum, mem_cumsum);
+
+          auto kernel =
+              vllm::moe::moe_align_block_size_global_mem_kernel<scalar_t>;
+          kernel<<<1, num_thread, 0, stream>>>(
+              topk_ids.data_ptr<scalar_t>(),
+              sorted_token_ids.data_ptr<int32_t>(),
+              experts_ids.data_ptr<int32_t>(),
+              num_tokens_post_pad.data_ptr<int32_t>(), num_experts, block_size,
+              topk_ids.numel(), tokens_cnts, cumsum);
+          cudaFree(tokens_cnts);
+          cudaFree(cumsum);
+        });
+  } else {
+    VLLM_DISPATCH_INTEGRAL_TYPES(
+        topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
+          // calc needed amount of shared mem for `tokens_cnts` and `cumsum`
+          // tensors
+          const int32_t num_thread = max((int32_t)num_experts, WARP_SIZE);
+          const int32_t shared_mem =
+              ((num_thread + 1) * num_experts + (num_experts + 1)) *
+              sizeof(int32_t);
+
+          // set dynamic shared mem
+          auto kernel = vllm::moe::moe_align_block_size_kernel<scalar_t>;
+          AT_CUDA_CHECK(VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(
+              (void*)kernel, shared_mem));
+          kernel<<<1, num_thread, shared_mem, stream>>>(
+              topk_ids.data_ptr<scalar_t>(),
+              sorted_token_ids.data_ptr<int32_t>(),
+              experts_ids.data_ptr<int32_t>(),
+              num_tokens_post_pad.data_ptr<int32_t>(), num_experts, block_size,
+              topk_ids.numel());
+        });
+  }
 }
 
 void moe_sum(torch::Tensor& input,   // [num_tokens, topk, hidden_size]

vllm/config.py

@@ -596,6 +596,12 @@ class ModelConfig:
         self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
                                           self.max_model_len)
 
+        if (self.hf_config.model_type == 'deepseek_v3'
+                and not self.enforce_eager):
+            logger.warning("CUDA graph is not supported for Deepseek V3 yet, "
+                           "fallback to the eager mode.")
+            self.enforce_eager = True
+
     def _verify_bnb_config(self) -> None:
         """
         The current version of bitsandbytes (0.44.0) with 8-bit models does not
@@ -712,8 +718,9 @@ class ModelConfig:
 
     def get_head_size(self) -> int:
         # TODO remove hard code
-        if hasattr(self.hf_text_config, "model_type"
-                   ) and self.hf_text_config.model_type == 'deepseek_v2':
+        if hasattr(self.hf_text_config,
+                   "model_type") and (self.hf_text_config.model_type
+                                      in ('deepseek_v2', 'deepseek_v3')):
             # FlashAttention supports only head_size 32, 64, 128, 256,
             # we need to pad head_size 192 to 256
             return 256

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -476,18 +476,29 @@ def fused_topk(
     return topk_weights, topk_ids
 
 
-# This is used by the Deepseek-V2 model
+# This is used by the Deepseek-V2 and Deepseek-V3 model
 def grouped_topk(hidden_states: torch.Tensor,
                  gating_output: torch.Tensor,
                  topk: int,
                  renormalize: bool,
                  num_expert_group: int = 0,
-                 topk_group: int = 0):
+                 topk_group: int = 0,
+                 scoring_func: str = "softmax",
+                 e_score_correction_bias: Optional[torch.Tensor] = None):
 
     assert hidden_states.shape[0] == gating_output.shape[0], (
         "Number of tokens mismatch")
 
-    scores = torch.softmax(gating_output, dim=-1)
+    if scoring_func == "softmax":
+        scores = torch.softmax(gating_output, dim=-1)
+    elif scoring_func == "sigmoid":
+        scores = gating_output.sigmoid()
+    else:
+        raise ValueError(f"Unsupported scoring function: {scoring_func}")
+
+    if e_score_correction_bias is not None:
+        scores.add_(e_score_correction_bias.unsqueeze(0))
+
     num_token = scores.shape[0]
     group_scores = scores.view(num_token, num_expert_group,
                                -1).max(dim=-1).values  # [n, n_group]

vllm/model_executor/layers/fused_moe/layer.py

@@ -73,16 +73,18 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         set_weight_attrs(w2_weight, extra_weight_attrs)
 
     def apply(
-            self,
-            layer: torch.nn.Module,
-            x: torch.Tensor,
-            router_logits: torch.Tensor,
-            top_k: int,
-            renormalize: bool,
-            use_grouped_topk: bool,
-            topk_group: Optional[int] = None,
-            num_expert_group: Optional[int] = None,
-            custom_routing_function: Optional[Callable] = None
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        router_logits: torch.Tensor,
+        top_k: int,
+        renormalize: bool,
+        use_grouped_topk: bool,
+        topk_group: Optional[int] = None,
+        num_expert_group: Optional[int] = None,
+        custom_routing_function: Optional[Callable] = None,
+        scoring_func: str = "softmax",
+        e_score_correction_bias: Optional[torch.Tensor] = None
     ) -> torch.Tensor:
         return self.forward(x=x,
                             layer=layer,
@@ -92,19 +94,23 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
                             use_grouped_topk=use_grouped_topk,
                             topk_group=topk_group,
                             num_expert_group=num_expert_group,
-                            custom_routing_function=custom_routing_function)
+                            custom_routing_function=custom_routing_function,
+                            scoring_func=scoring_func,
+                            e_score_correction_bias=e_score_correction_bias)
 
     def forward_cuda(
-            self,
-            layer: torch.nn.Module,
-            x: torch.Tensor,
-            use_grouped_topk: bool,
-            top_k: int,
-            router_logits: torch.Tensor,
-            renormalize: bool,
-            topk_group: Optional[int] = None,
-            num_expert_group: Optional[int] = None,
-            custom_routing_function: Optional[Callable] = None
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        use_grouped_topk: bool,
+        top_k: int,
+        router_logits: torch.Tensor,
+        renormalize: bool,
+        topk_group: Optional[int] = None,
+        num_expert_group: Optional[int] = None,
+        custom_routing_function: Optional[Callable] = None,
+        scoring_func: str = "softmax",
+        e_score_correction_bias: Optional[torch.Tensor] = None
     ) -> torch.Tensor:
         topk_weights, topk_ids = FusedMoE.select_experts(
             hidden_states=x,
@@ -114,7 +120,9 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
             renormalize=renormalize,
             topk_group=topk_group,
             num_expert_group=num_expert_group,
-            custom_routing_function=custom_routing_function)
+            custom_routing_function=custom_routing_function,
+            scoring_func=scoring_func,
+            e_score_correction_bias=e_score_correction_bias)
 
         return fused_experts(hidden_states=x,
                              w1=layer.w13_weight,
@@ -128,21 +136,29 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
             "The CPU backend currently does not support MoE.")
 
     def forward_tpu(
-            self,
-            layer: torch.nn.Module,
-            x: torch.Tensor,
-            use_grouped_topk: bool,
-            top_k: int,
-            router_logits: torch.Tensor,
-            renormalize: bool,
-            topk_group: Optional[int] = None,
-            num_expert_group: Optional[int] = None,
-            custom_routing_function: Optional[Callable] = None
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        use_grouped_topk: bool,
+        top_k: int,
+        router_logits: torch.Tensor,
+        renormalize: bool,
+        topk_group: Optional[int] = None,
+        num_expert_group: Optional[int] = None,
+        custom_routing_function: Optional[Callable] = None,
+        scoring_func: str = "softmax",
+        e_score_correction_bias: Optional[torch.Tensor] = None
     ) -> torch.Tensor:
         assert not use_grouped_topk
         assert num_expert_group is None
         assert topk_group is None
         assert custom_routing_function is None
+        if scoring_func != "softmax":
+            raise NotImplementedError(
+                "Only softmax scoring function is supported for TPU.")
+        if e_score_correction_bias is not None:
+            raise NotImplementedError(
+                "Expert score correction bias is not supported for TPU.")
         return fused_moe_pallas(hidden_states=x,
                                 w1=layer.w13_weight,
                                 w2=layer.w2_weight,
@@ -156,7 +172,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
 class FusedMoE(torch.nn.Module):
     """FusedMoE layer for MoE models.
 
-    This layer contains both MergedColumnParallel weights (gate_up_proj / 
+    This layer contains both MergedColumnParallel weights (gate_up_proj /
     w13) and RowParallelLinear weights (down_proj/ w2).
 
     Note: Mixtral uses w1, w2, and w3 for gate, up, and down_proj. We
@@ -190,6 +206,8 @@ class FusedMoE(torch.nn.Module):
         tp_size: Optional[int] = None,
         prefix: str = "",
         custom_routing_function: Optional[Callable] = None,
+        scoring_func: str = "softmax",
+        e_score_correction_bias: Optional[torch.Tensor] = None,
     ):
         super().__init__()
 
@@ -210,6 +228,12 @@ class FusedMoE(torch.nn.Module):
         self.num_expert_group = num_expert_group
         self.topk_group = topk_group
         self.custom_routing_function = custom_routing_function
+        self.scoring_func = scoring_func
+        self.e_score_correction_bias = e_score_correction_bias
+
+        if self.scoring_func != "softmax" and not self.use_grouped_topk:
+            raise ValueError("Only softmax scoring function is supported for "
+                             "non-grouped topk.")
 
         if quant_config is None:
             self.quant_method: Optional[QuantizeMethodBase] = (
@@ -446,7 +470,9 @@ class FusedMoE(torch.nn.Module):
                        renormalize: bool,
                        topk_group: Optional[int] = None,
                        num_expert_group: Optional[int] = None,
-                       custom_routing_function: Optional[Callable] = None):
+                       custom_routing_function: Optional[Callable] = None,
+                       scoring_func: str = "softmax",
+                       e_score_correction_bias: Optional[torch.Tensor] = None):
         from vllm.model_executor.layers.fused_moe.fused_moe import (
             fused_topk, grouped_topk)
 
@@ -460,7 +486,9 @@ class FusedMoE(torch.nn.Module):
                 topk=top_k,
                 renormalize=renormalize,
                 num_expert_group=num_expert_group,
-                topk_group=topk_group)
+                topk_group=topk_group,
+                scoring_func=scoring_func,
+                e_score_correction_bias=e_score_correction_bias)
         elif custom_routing_function is None:
             topk_weights, topk_ids = fused_topk(hidden_states=hidden_states,
                                                 gating_output=router_logits,
@@ -489,7 +517,9 @@ class FusedMoE(torch.nn.Module):
             use_grouped_topk=self.use_grouped_topk,
             topk_group=self.topk_group,
             num_expert_group=self.num_expert_group,
-            custom_routing_function=self.custom_routing_function)
+            custom_routing_function=self.custom_routing_function,
+            scoring_func=self.scoring_func,
+            e_score_correction_bias=self.e_score_correction_bias)
 
         if self.reduce_results and self.tp_size > 1:
             final_hidden_states = tensor_model_parallel_all_reduce(

vllm/model_executor/layers/quantization/fp8.py

@@ -605,6 +605,8 @@ class Fp8MoEMethod(FusedMoEMethodBase):
         topk_group: Optional[int] = None,
         num_expert_group: Optional[int] = None,
         custom_routing_function: Optional[Callable] = None,
+        scoring_func: str = "softmax",
+        e_score_correction_bias: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
 
         from vllm.model_executor.layers.fused_moe import fused_experts
@@ -617,7 +619,10 @@ class Fp8MoEMethod(FusedMoEMethodBase):
             renormalize=renormalize,
             topk_group=topk_group,
             num_expert_group=num_expert_group,
-            custom_routing_function=custom_routing_function)
+            custom_routing_function=custom_routing_function,
+            scoring_func=scoring_func,
+            e_score_correction_bias=e_score_correction_bias,
+        )
 
         return fused_experts(
             x,

vllm/model_executor/models/deepseek_v3.py

+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
+# Copyright 2023 The vLLM team.
+# Copyright 2023 DeepSeek-AI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only DeepseekV3 model."""
+from typing import Any, Dict, Iterable, List, Optional, Set, Tuple, Union
+
+import torch
+from torch import nn
+from transformers import PretrainedConfig
+
+from vllm.attention import Attention, AttentionMetadata
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import (get_pp_group,
+                              get_tensor_model_parallel_world_size,
+                              tensor_model_parallel_all_reduce)
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               MergedColumnParallelLinear,
+                                               ReplicatedLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    ParallelLMHead, VocabParallelEmbedding)
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+
+from .interfaces import SupportsPP
+from .utils import (PPMissingLayer, is_pp_missing_parameter,
+                    make_empty_intermediate_tensors_factory, make_layers,
+                    maybe_prefix)
+
+
+class DeepseekV3MLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = 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
+
+
+class DeepseekV3MoE(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.n_shared_experts = config.n_shared_experts
+        self.routed_scaling_factor = config.routed_scaling_factor
+        if self.tp_size > config.n_routed_experts:
+            raise ValueError(
+                f"Tensor parallel size {self.tp_size} is greater than "
+                f"the number of experts {config.n_routed_experts}.")
+
+        if config.hidden_act != "silu":
+            raise ValueError(f"Unsupported activation: {config.hidden_act}. "
+                             "Only silu is supported for now.")
+
+        self.gate = ReplicatedLinear(config.hidden_size,
+                                     config.n_routed_experts,
+                                     bias=False,
+                                     quant_config=None,
+                                     prefix=f"{prefix}.gate")
+        if config.topk_method == "noaux_tc":
+            self.gate.e_score_correction_bias = nn.Parameter(
+                torch.empty(config.n_routed_experts))
+        else:
+            self.gate.e_score_correction_bias = None
+
+        self.experts = FusedMoE(
+            num_experts=config.n_routed_experts,
+            top_k=config.num_experts_per_tok,
+            hidden_size=config.hidden_size,
+            intermediate_size=config.moe_intermediate_size,
+            reduce_results=False,
+            renormalize=config.norm_topk_prob,
+            quant_config=quant_config,
+            use_grouped_topk=True,
+            num_expert_group=config.n_group,
+            topk_group=config.topk_group,
+            prefix=f"{prefix}.experts",
+            scoring_func=config.scoring_func,
+            e_score_correction_bias=self.gate.e_score_correction_bias)
+
+        if config.n_shared_experts is not None:
+            intermediate_size = (config.moe_intermediate_size *
+                                 config.n_shared_experts)
+            self.shared_experts = DeepseekV3MLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+                reduce_results=False,
+            )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        num_tokens, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        if self.n_shared_experts is not None:
+            shared_output = self.shared_experts(hidden_states)
+        # router_logits: (num_tokens, n_experts)
+        router_logits, _ = self.gate(hidden_states)
+        final_hidden_states = self.experts(
+            hidden_states=hidden_states,
+            router_logits=router_logits) * self.routed_scaling_factor
+        if shared_output is not None:
+            final_hidden_states = final_hidden_states + shared_output
+        if self.tp_size > 1:
+            final_hidden_states = tensor_model_parallel_all_reduce(
+                final_hidden_states)
+
+        return final_hidden_states.view(num_tokens, hidden_dim)
+
+
+def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
+    import math
+    if scale <= 1:
+        return 1.0
+    return 0.1 * mscale * math.log(scale) + 1.0
+
+
+class DeepseekV3Attention(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        hidden_size: int,
+        num_heads: int,
+        qk_nope_head_dim: int,
+        qk_rope_head_dim: int,
+        v_head_dim: int,
+        q_lora_rank: int,
+        kv_lora_rank: int,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 8192,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.q_lora_rank = q_lora_rank
+        self.kv_lora_rank = kv_lora_rank
+        self.num_heads = num_heads
+        tp_size = get_tensor_model_parallel_world_size()
+        assert num_heads % tp_size == 0
+        self.num_local_heads = num_heads // tp_size
+        self.scaling = self.qk_head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        if self.q_lora_rank is not None:
+            self.q_a_proj = ReplicatedLinear(self.hidden_size,
+                                             self.q_lora_rank,
+                                             bias=False,
+                                             quant_config=quant_config,
+                                             prefix=f"{prefix}.q_a_proj")
+            self.q_a_layernorm = RMSNorm(self.q_lora_rank,
+                                         eps=config.rms_norm_eps)
+            self.q_b_proj = ColumnParallelLinear(q_lora_rank,
+                                                 self.num_heads *
+                                                 self.qk_head_dim,
+                                                 bias=False,
+                                                 quant_config=quant_config,
+                                                 prefix=f"{prefix}.q_b_proj")
+        else:
+            self.q_proj = ColumnParallelLinear(self.hidden_size,
+                                               self.num_heads *
+                                               self.qk_head_dim,
+                                               bias=False,
+                                               quant_config=quant_config,
+                                               prefix=f"{prefix}.q_proj")
+
+        self.kv_a_proj_with_mqa = ReplicatedLinear(
+            self.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.kv_a_proj_with_mqa")
+        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank,
+                                      eps=config.rms_norm_eps)
+        self.kv_b_proj = ColumnParallelLinear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.kv_b_proj")
+        # O projection.
+        self.o_proj = RowParallelLinear(self.num_heads * self.v_head_dim,
+                                        self.hidden_size,
+                                        bias=False,
+                                        quant_config=quant_config,
+                                        prefix=f"{prefix}.o_proj")
+        rope_scaling["rope_type"] = 'deepseek_yarn'
+        self.rotary_emb = get_rope(qk_rope_head_dim,
+                                   rotary_dim=qk_rope_head_dim,
+                                   max_position=max_position_embeddings,
+                                   base=rope_theta,
+                                   rope_scaling=rope_scaling,
+                                   is_neox_style=False)
+
+        if rope_scaling:
+            mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
+            scaling_factor = rope_scaling["factor"]
+            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
+            self.scaling = self.scaling * mscale * mscale
+
+        # self.attn = Attention(self.num_heads,
+        #                       self.qk_head_dim,
+        #                       self.scaling,
+        #                       num_kv_heads=self.num_heads)
+
+        # TODO, support head_size 192
+        self.attn = Attention(self.num_local_heads,
+                              256,
+                              self.scaling,
+                              num_kv_heads=self.num_local_heads,
+                              cache_config=cache_config,
+                              quant_config=quant_config,
+                              prefix=f"{prefix}.attn")
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        if self.q_lora_rank is not None:
+            q = self.q_a_proj(hidden_states)[0]
+            q = self.q_a_layernorm(q)
+            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads,
+                                         self.qk_head_dim)
+        else:
+            q = self.q_proj(hidden_states)[0].view(-1, self.num_local_heads,
+                                                   self.qk_head_dim)
+        q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim],
+                               dim=-1)
+        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
+        kv_a, _ = latent_cache.split(
+            [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+        latent_cache = latent_cache.unsqueeze(1)
+        kv_a = self.kv_a_layernorm(kv_a.contiguous())
+        kv = self.kv_b_proj(kv_a)[0]
+        kv = kv.view(-1, self.num_local_heads,
+                     self.qk_nope_head_dim + self.v_head_dim)
+        k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+        k_pe = latent_cache[:, :, self.kv_lora_rank:]
+        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
+        q[..., self.qk_nope_head_dim:] = q_pe
+        k = torch.empty_like(q)
+        k[..., :self.qk_nope_head_dim] = k_nope
+        k[..., self.qk_nope_head_dim:] = k_pe
+        q = torch.nn.functional.pad(q, [0, 256 - self.qk_head_dim],
+                                    value=0).view(-1,
+                                                  self.num_local_heads * 256)
+        k = torch.nn.functional.pad(k, [0, 256 - self.qk_head_dim],
+                                    value=0).view(-1,
+                                                  self.num_local_heads * 256)
+        v = torch.nn.functional.pad(v, [0, 256 - self.v_head_dim],
+                                    value=0).view(-1,
+                                                  self.num_local_heads * 256)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
+        attn_output = attn_output.view(
+            -1, self.num_local_heads, 256)[..., :self.v_head_dim].reshape(
+                -1, self.num_local_heads * self.v_head_dim)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class DeepseekV3DecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        prefix: str,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        # DecoderLayers are created with `make_layers` which passes the prefix
+        # with the layer's index.
+        layer_idx = int(prefix.split(sep='.')[-1])
+        self.self_attn = DeepseekV3Attention(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            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
+            if hasattr(config, "q_lora_rank") else None,
+            kv_lora_rank=config.kv_lora_rank,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.self_attn",
+        )
+        if (config.n_routed_experts is not None
+                and layer_idx >= config.first_k_dense_replace
+                and layer_idx % config.moe_layer_freq == 0):
+            self.mlp = DeepseekV3MoE(
+                config=config,
+                quant_config=quant_config,
+                prefix=f"{prefix}.mlp",
+            )
+        else:
+            self.mlp = DeepseekV3MLP(
+                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,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+        residual: Optional[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,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(
+            hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+# TODO(simon): check whether we support torch compile for Deepseek V3
+# @support_torch_compile
+class DeepseekV3Model(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
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        if get_pp_group().is_first_rank:
+            self.embed_tokens = VocabParallelEmbedding(
+                config.vocab_size,
+                config.hidden_size,
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+
+        self.start_layer, self.end_layer, self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: DeepseekV3DecoderLayer(
+                config,
+                prefix,
+                cache_config=cache_config,
+                quant_config=quant_config,
+            ),
+            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))
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+            residual = None
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.layers[i]
+            hidden_states, residual = layer(positions, hidden_states,
+                                            kv_caches[i - self.start_layer],
+                                            attn_metadata, residual)
+
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({
+                "hidden_states": hidden_states,
+                "residual": residual
+            })
+
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class DeepseekV3ForCausalLM(nn.Module, SupportsPP):
+
+    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
+        self.quant_config = quant_config
+        self.model = DeepseekV3Model(vllm_config=vllm_config,
+                                     prefix=maybe_prefix(prefix, "model"))
+        self.lm_head = ParallelLMHead(config.vocab_size,
+                                      config.hidden_size,
+                                      quant_config=quant_config)
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+        self.sampler = get_sampler()
+        self.make_empty_intermediate_tensors = (
+            self.model.make_empty_intermediate_tensors)
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        hidden_states = self.model(input_ids, positions, kv_caches,
+                                   attn_metadata, intermediate_tensors,
+                                   inputs_embeds)
+        return hidden_states
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[torch.Tensor]:
+        logits = self.logits_processor(self.lm_head, hidden_states,
+                                       sampling_metadata)
+        return logits
+
+    def sample(
+        self,
+        logits: Optional[torch.Tensor],
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[SamplerOutput]:
+        next_tokens = self.sampler(logits, sampling_metadata)
+        return next_tokens
+
+    def make_empty_intermediate_tensors(
+            self, batch_size: int, dtype: torch.dtype,
+            device: torch.device) -> IntermediateTensors:
+        return IntermediateTensors({
+            "hidden_states":
+            torch.zeros((batch_size, self.config.hidden_size),
+                        dtype=dtype,
+                        device=device),
+            "residual":
+            torch.zeros((batch_size, self.config.hidden_size),
+                        dtype=dtype,
+                        device=device),
+        })
+
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+
+        # Params for weights, fp8 weight scales, fp8 activation scales
+        # (param_name, weight_name, expert_id, shard_id)
+        expert_params_mapping = FusedMoE.make_expert_params_mapping(
+            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)
+
+        params_dict = dict(self.named_parameters())
+        loaded_params: Set[str] = set()
+        for name, loaded_weight in weights:
+            if "rotary_emb.inv_freq" in name:
+                continue
+
+            # TODO(simon): support nextn predict layers
+            if self.config.num_nextn_predict_layers > 0:
+                assert self.config.num_nextn_predict_layers == 1
+                layer_idx = self.config.num_hidden_layers
+                if name.startswith(f"model.layers.{layer_idx}"):
+                    continue
+
+            for (param_name, weight_name, shard_id) in stacked_params_mapping:
+                # Skip non-stacked layers and experts (experts handled below).
+                if weight_name not in name:
+                    continue
+                # We have mlp.experts[0].gate_proj in the checkpoint.
+                # Since we handle the experts below in expert_params_mapping,
+                # we need to skip here BEFORE we update the name, otherwise
+                # name will be updated to mlp.experts[0].gate_up_proj, which
+                # will then be updated below in expert_params_mapping
+                # for mlp.experts[0].gate_gate_up_proj, which breaks load.
+                if (("mlp.experts." in name) and name not in params_dict):
+                    continue
+                name = name.replace(weight_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+
+                if is_pp_missing_parameter(name, self):
+                    continue
+
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                for mapping in expert_params_mapping:
+                    param_name, weight_name, expert_id, shard_id = mapping
+                    if weight_name not in name:
+                        continue
+                    name = name.replace(weight_name, param_name)
+
+                    if is_pp_missing_parameter(name, self):
+                        continue
+
+                    param = params_dict[name]
+                    weight_loader = param.weight_loader
+                    weight_loader(param,
+                                  loaded_weight,
+                                  name,
+                                  shard_id=shard_id,
+                                  expert_id=expert_id)
+                    break
+                else:
+                    # Skip loading extra bias for GPTQ models.
+                    if name.endswith(".bias") and name not in params_dict:
+                        continue
+
+                    if is_pp_missing_parameter(name, self):
+                        continue
+
+                    if name not in params_dict:
+                        for key in params_dict:
+                            print(key)
+                    param = params_dict[name]
+                    weight_loader = getattr(param, "weight_loader",
+                                            default_weight_loader)
+                    weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params

vllm/model_executor/models/registry.py

@@ -45,6 +45,7 @@ _TEXT_GENERATION_MODELS = {
     "DeciLMForCausalLM": ("decilm", "DeciLMForCausalLM"),
     "DeepseekForCausalLM": ("deepseek", "DeepseekForCausalLM"),
     "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
+    "DeepseekV3ForCausalLM": ("deepseek_v3", "DeepseekV3ForCausalLM"),
     "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
     "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
     "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),