ref(upstream): 搬运 xllm ilu kernel+layer 完整源码 — 2089行 14个API声明

来源: Deep-Spark/xllm core/kernels/ilu/ + core/layers/ilu/ ixformer.h: 14个ixformer::infer API完整声明 kernel wrappers: activation(32) attention(162) fused_moe(99) group_gemm(39) matmul(73) norm(50) rope(31) + headers layer dispatch: fused_moe.cpp(797行) attention.cpp(189行) + headers 覆盖状态 (ix_moe_bridge.cpp vs ixformer.h 14个API): 已覆盖 13/14: silu_and_mul, rms_norm, residual_rms_norm, ixformer_linear, ixformer_linear_ex, topk_softmax, moe_compute_token_index, moe_expand_input, moe_w16a16_group_gemm, moe_output_reduce_sum, xllm_paged_attention, xllm_reshape_and_cache, xllm_rotary_embedding 缺失 1/14: ixinfer_flash_attn_unpad_with_block_tables dlopen 调用链验证: 12个 prebuilt .so → 9个 qwen3_5.py + 2个 paged_attn.py + 1个 block_major_kv_cache.py 辅助模块: bi100_env, bi100_profile, gdn_prefix, block_major_kv_cache 全部到位
2026-08-11 04:41:29 +00:00
parent 6cdf2ec87b
commit 32ee28122e
16 changed files with 69 additions and 36 deletions
--- a/upstream_ref/xllm_latest/core/kernels/ilu/CMakeLists.txt
+++ b/upstream_ref/xllm_latest/core/kernels/ilu/CMakeLists.txt
@@ -0,0 +1,28 @@
+include(cc_library)
+set(CMAKE_CUDA_ARCHITECTURES ivcore11)
+file(GLOB_RECURSE ILU_HEADER_FILES
+  "${CMAKE_CURRENT_LIST_DIR}/*.h"
+)
+
+file(GLOB_RECURSE ILU_SOURCE_FILES
+  "${CMAKE_CURRENT_LIST_DIR}/*.cpp"
+  "${CMAKE_CURRENT_LIST_DIR}/*.cu"
+)
+
+find_package(Python3 REQUIRED COMPONENTS Interpreter Development)
+
+cc_library(
+  NAME
+    ilu_kernels
+  HDRS
+    ${ILU_HEADER_FILES}
+  SRCS
+    ${ILU_SOURCE_FILES}
+  DEPS
+    torch
+    :util
+    ixformer_kernels
+    ixformer
+    ${Python3_LIBRARIES}
+    cuinfer
+)
--- a/upstream_ref/xllm_latest/core/kernels/ilu/activation.cpp
+++ b/upstream_ref/xllm_latest/core/kernels/ilu/activation.cpp
@@ -1,4 +1,4 @@
-/* Copyright 2025-2026 The xLLM Authors.
+/* Copyright 2025 The xLLM Authors. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
--- a/upstream_ref/xllm_latest/core/kernels/ilu/attention.cpp
+++ b/upstream_ref/xllm_latest/core/kernels/ilu/attention.cpp
@@ -1,5 +1,5 @@

-/* Copyright 2025-2026 The xLLM Authors.
+/* Copyright 2025 The xLLM Authors. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
--- a/upstream_ref/xllm_latest/core/kernels/ilu/fused_moe.cpp
+++ b/upstream_ref/xllm_latest/core/kernels/ilu/fused_moe.cpp
@@ -1,4 +1,4 @@
-/* Copyright 2025-2026 The xLLM Authors.
+/* Copyright 2025 The xLLM Authors. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
--- a/upstream_ref/xllm_latest/core/kernels/ilu/group_gemm.cpp
+++ b/upstream_ref/xllm_latest/core/kernels/ilu/group_gemm.cpp
@@ -1,4 +1,4 @@
-/* Copyright 2025-2026 The xLLM Authors.
+/* Copyright 2026 The xLLM Authors. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
--- a/upstream_ref/xllm_latest/core/kernels/ilu/ilu_ops_api.h
+++ b/upstream_ref/xllm_latest/core/kernels/ilu/ilu_ops_api.h
@@ -1,4 +1,4 @@
-/* Copyright 2025-2026 The xLLM Authors.
+/* Copyright 2025 The xLLM Authors. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
--- a/upstream_ref/xllm_latest/core/kernels/ilu/ixformer.h
+++ b/upstream_ref/xllm_latest/core/kernels/ilu/ixformer.h
@@ -1,4 +1,4 @@
-/* Copyright 2025-2026 The xLLM Authors.
+/* Copyright 2025 The xLLM Authors. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
--- a/upstream_ref/xllm_latest/core/kernels/ilu/matmul.cpp
+++ b/upstream_ref/xllm_latest/core/kernels/ilu/matmul.cpp
@@ -1,4 +1,4 @@
-/* Copyright 2025-2026 The xLLM Authors.
+/* Copyright 2025 The xLLM Authors. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
--- a/upstream_ref/xllm_latest/core/kernels/ilu/norm.cpp
+++ b/upstream_ref/xllm_latest/core/kernels/ilu/norm.cpp
@@ -1,4 +1,4 @@
-/* Copyright 2025-2026 The xLLM Authors.
+/* Copyright 2025 The xLLM Authors. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
--- a/upstream_ref/xllm_latest/core/kernels/ilu/rope.cpp
+++ b/upstream_ref/xllm_latest/core/kernels/ilu/rope.cpp
@@ -1,4 +1,4 @@
-/* Copyright 2025-2026 The xLLM Authors.
+/* Copyright 2025 The xLLM Authors. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
--- a/upstream_ref/xllm_latest/core/kernels/ilu/utils.h
+++ b/upstream_ref/xllm_latest/core/kernels/ilu/utils.h
@@ -1,4 +1,4 @@
-/* Copyright 2025-2026 The xLLM Authors.
+/* Copyright 2025 The xLLM Authors. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
--- a/upstream_ref/xllm_latest/core/layers/ilu/CMakeLists.txt
+++ b/upstream_ref/xllm_latest/core/layers/ilu/CMakeLists.txt
@@ -0,0 +1,14 @@
+include(cc_library)
+
+cc_library(
+  NAME
+    ilu_layers
+  HDRS
+    attention.h
+    fused_moe.h
+  SRCS
+    attention.cpp
+    fused_moe.cpp
+  DEPS
+    :common_layers
+)
--- a/upstream_ref/xllm_latest/core/layers/ilu/attention.cpp
+++ b/upstream_ref/xllm_latest/core/layers/ilu/attention.cpp
@@ -1,4 +1,4 @@
-/* Copyright 2025-2026 The xLLM Authors.
+/* Copyright 2025 The xLLM Authors. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
--- a/upstream_ref/xllm_latest/core/layers/ilu/attention.h
+++ b/upstream_ref/xllm_latest/core/layers/ilu/attention.h
@@ -1,4 +1,4 @@
-/* Copyright 2025-2026 The xLLM Authors.
+/* Copyright 2025 The xLLM Authors. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
--- a/upstream_ref/xllm_latest/core/layers/ilu/fused_moe.cpp
+++ b/upstream_ref/xllm_latest/core/layers/ilu/fused_moe.cpp
@@ -1,4 +1,4 @@
-/* Copyright 2025-2026 The xLLM Authors.
+/* Copyright 2026 The xLLM Authors. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
@@ -20,9 +20,6 @@ limitations under the License.
 #include <iomanip>

 #include "common/global_flags.h"
-#include "core/framework/config/eplb_config.h"
-#include "core/framework/config/scheduler_config.h"
-#include "core/framework/config/speculative_config.h"
 #include "framework/parallel_state/parallel_state.h"
 #include "kernels/ops_api.h"
 #include "layers/common/dp_utils.h"
@@ -89,9 +86,7 @@ FusedMoEImpl::FusedMoEImpl(const ModelArgs& model_args,
  }

  // Deep EP initialization check
-  enable_deep_ep_ =
-      ::xllm::EPLBConfig::get_instance().expert_parallel_degree() == 2 &&
-      ep_size > 1;
+  enable_deep_ep_ = FLAGS_expert_parallel_degree == 2 && ep_size > 1;
  if (enable_deep_ep_) {
    // for now, we only implement the deep ep for decode stage.
    // so we will assume the max_token_num is limited to max_batch_size * (1+K)
@@ -108,20 +103,16 @@ FusedMoEImpl::FusedMoEImpl(const ModelArgs& model_args,
    torch::ScalarType combine_dtype = options_.dtype().toScalarType();
    int64_t combine_token_size = hidden_size_ * get_dtype_size(combine_dtype);
    // Ensure calculation base is at least ep_size
-    int64_t effective_seqs = std::max(
-        (int64_t)::xllm::SchedulerConfig::get_instance().max_seqs_per_batch(),
-        (int64_t)ep_size);
-    // NOTE: ::xllm::SchedulerConfig::get_instance().max_seqs_per_batch()
-    // represents the maximum total batch size, regardless of the dp size. To
-    // ensure robust scheduling and account for the worst-case scenario, we must
-    // guarantee that each rank is capable of handling the maximum possible
-    // number of tokens. Therefore, we define max_num_tokens_per_rank as the
-    // full maximum value, without dividing by either the rank count or the dp
-    // size.
+    int64_t effective_seqs =
+        std::max((int64_t)FLAGS_max_seqs_per_batch, (int64_t)ep_size);
+    // NOTE: FLAGS_max_seqs_per_batch represents the maximum total batch size,
+    // regardless of the dp size. To ensure robust scheduling and account
+    // for the worst-case scenario, we must guarantee that each rank is capable
+    // of handling the maximum possible number of tokens. Therefore, we define
+    // max_num_tokens_per_rank as the full maximum value, without dividing by
+    // either the rank count or the dp size.
    int64_t max_num_tokens_per_rank =
-        (1 +
-         ::xllm::SpeculativeConfig::get_instance().num_speculative_tokens()) *
-        effective_seqs * topk_;
+        (1 + FLAGS_num_speculative_tokens) * effective_seqs * topk_;

    // make sure that all layers share the same deep ep instance
    //  so that the memory footprint is minimized
@@ -723,8 +714,8 @@ torch::Tensor FusedMoEImpl::forward(const torch::Tensor& hidden_states,
                                    const ModelInputParams& input_params) {
  // we only support all2all communication for decode stage for now
  bool enable_all2all_communication =
-      enable_deep_ep_ && std::all_of(input_params.parallel.dp_is_decode.begin(),
-                                     input_params.parallel.dp_is_decode.end(),
+      enable_deep_ep_ && std::all_of(input_params.dp_is_decode.begin(),
+                                     input_params.dp_is_decode.end(),
                                     [](int32_t val) { return val == 1; });

  bool is_dp_ep_parallel =
@@ -739,7 +730,7 @@ torch::Tensor FusedMoEImpl::forward(const torch::Tensor& hidden_states,
  if (need_gather_and_slice) {
    input = parallel_state::gather(input,
                                   parallel_args_.dp_local_process_group_,
-                                   input_params.parallel.dp_global_token_nums);
+                                   input_params.dp_global_token_nums);
  }
  // MoE Gate
  auto router_logits = gate_(input);
--- a/upstream_ref/xllm_latest/core/layers/ilu/fused_moe.h
+++ b/upstream_ref/xllm_latest/core/layers/ilu/fused_moe.h
@@ -1,4 +1,4 @@
-/* Copyright 2025-2026 The xLLM Authors.
+/* Copyright 2025 The xLLM Authors. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.