Merge remote-tracking branch 'origin/main' into kv-xfer-updates

Author: wseaton

.buildkite/test-pipeline.yaml

@@ -41,6 +41,16 @@ steps:
   # TODO: add `--strict` once warnings in docstrings are fixed
   - mkdocs build
 
+- label: Pytorch Nightly Dependency Override Check # 2min
+  # if this test fails, it means the nightly torch version is not compatible with some
+  # of the dependencies. Please check the error message and add the package to whitelist
+  # in /vllm/tools/generate_nightly_torch_test.py
+  soft_fail: true
+  source_file_dependencies:
+  - requirements/nightly_torch_test.txt
+  commands:
+  - bash standalone_tests/pytorch_nightly_dependency.sh
+
 - label: Async Engine, Inputs, Utils, Worker Test # 24min
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
@@ -168,6 +178,23 @@ steps:
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
   - popd
 
+- label: EPLB Algorithm Test
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/distributed/eplb
+  - tests/distributed/test_eplb_algo.py
+  commands:
+  - pytest -v -s distributed/test_eplb_algo.py
+
+- label: EPLB Execution Test # 5min
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 4
+  source_file_dependencies:
+  - vllm/distributed/eplb
+  - tests/distributed/test_eplb_execute.py
+  commands:
+  - pytest -v -s distributed/test_eplb_execute.py
+
 - label: Metrics, Tracing Test # 10min
   mirror_hardwares: [amdexperimental, amdproduction]
   num_gpus: 2
@@ -750,7 +777,7 @@ steps:
     - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models.txt
 
 - label: Weight Loading Multiple GPU Test - Large Models # optional
-  mirror_hardwares: [amdexperimental] 
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   gpu: a100

.pre-commit-config.yaml

@@ -53,6 +53,11 @@ repos:
       files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:
+  - id: format-torch-nightly-test
+    name: reformat nightly_torch_test.txt to be in sync with test.in
+    language: python
+    entry: python tools/generate_nightly_torch_test.py
+    files: ^requirements/test\.(in|txt)$
   - id: mypy-local
     name: Run mypy for local Python installation
     entry: tools/mypy.sh 0 "local"

CMakeLists.txt

@@ -513,6 +513,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       CUDA_ARCHS "${FP4_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4=1")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
     message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}")
   else()
     message(STATUS "Not building NVFP4 as no compatible archs were found.")
@@ -547,8 +548,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # if it's possible to compile MoE kernels that use its output.
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu"
-             "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -566,6 +566,16 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
+  # moe_data.cu is used by all CUTLASS MoE kernels.
+  cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${CUTLASS_MOE_DATA_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+  endif() 
+
   #
   # Machete kernels
 
@@ -638,6 +648,14 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 # if CUDA endif
 endif()
 
+if (VLLM_GPU_LANG STREQUAL "HIP")
+  # Add QuickReduce kernels
+  list(APPEND VLLM_EXT_SRC
+    "csrc/custom_quickreduce.cu"
+  )
+# if ROCM endif
+endif()
+
 message(STATUS "Enabling C extension.")
 define_gpu_extension_target(
   _C

csrc/custom_quickreduce.cu

@@ -0,0 +1,114 @@
+#include <ATen/cuda/Exceptions.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <c10/cuda/CUDAStream.h>
+#include <torch/all.h>
+
+#ifdef USE_ROCM
+
+  #include "quickreduce/quick_reduce.h"
+
+quickreduce::fptr_t init_custom_qr(int64_t rank, int64_t world_size,
+                                   std::optional<int64_t> qr_max_size) {
+  if (world_size > 8)
+    throw std::invalid_argument("world size > 8 is not supported");
+  if (world_size == 6)
+    throw std::invalid_argument("world size == 6 is not supported");
+  if (world_size % 2 != 0)
+    throw std::invalid_argument("Odd num gpus is not supported for now");
+  if (rank < 0 || rank >= world_size)
+    throw std::invalid_argument("invalid rank passed in");
+  quickreduce::DeviceComms* fptr = new quickreduce::DeviceComms();
+  fptr->init(world_size, rank, qr_max_size);
+  return (quickreduce::fptr_t)fptr;
+}
+
+void qr_destroy(quickreduce::fptr_t _fa) {
+  if (_fa) {
+    auto fa = reinterpret_cast<quickreduce::DeviceComms*>(_fa);
+    fa->destroy();
+    delete fa;
+  }
+}
+
+torch::Tensor qr_get_handle(quickreduce::fptr_t _fa) {
+  auto fa = reinterpret_cast<quickreduce::DeviceComms*>(_fa);
+  hipIpcMemHandle_t handle = fa->get_handle();
+  auto options =
+      torch::TensorOptions().dtype(torch::kUInt8).device(torch::kCPU);
+  auto data_handle =
+      torch::empty({static_cast<int64_t>(sizeof(hipIpcMemHandle_t))}, options);
+  std::memcpy(data_handle.data_ptr(), &handle, sizeof(hipIpcMemHandle_t));
+  return data_handle;
+}
+
+void qr_open_handles(quickreduce::fptr_t _fa,
+                     const std::vector<torch::Tensor>& handles) {
+  auto fa = reinterpret_cast<quickreduce::DeviceComms*>(_fa);
+  std::vector<hipIpcMemHandle_t> ipc_handles;
+  ipc_handles.reserve(handles.size());
+  for (auto& handle : handles) {
+    // Ensure the tensor is on the same device as the current device.
+    hipIpcMemHandle_t ipc_handle;
+    std::memcpy(&ipc_handle, handle.data_ptr(), sizeof(hipIpcMemHandle_t));
+    ipc_handles.push_back(ipc_handle);
+  }
+  fa->open_ipc_handles(ipc_handles);
+}
+
+void qr_all_reduce(quickreduce::fptr_t _fa, torch::Tensor& inp,
+                   torch::Tensor& out, int64_t quant_level, bool cast_bf2half) {
+  auto fa = reinterpret_cast<quickreduce::DeviceComms*>(_fa);
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(inp));
+  auto stream = at::cuda::getCurrentHIPStreamMasqueradingAsCUDA();
+
+  TORCH_CHECK_EQ(inp.scalar_type(), out.scalar_type());
+  TORCH_CHECK_EQ(inp.numel(), out.numel());
+  TORCH_CHECK_LE(out.numel(), fa->kMaxProblemSize);
+  if (out.scalar_type() == at::ScalarType::Half) {
+    fa->allreduce<half, false>(reinterpret_cast<half*>(inp.data_ptr()),
+                               reinterpret_cast<half*>(out.data_ptr()),
+                               out.numel(), quant_level, stream);
+  } else if (out.scalar_type() == at::ScalarType::BFloat16) {
+    if (cast_bf2half) {
+      fa->allreduce<half, true>(reinterpret_cast<half*>(inp.data_ptr()),
+                                reinterpret_cast<half*>(out.data_ptr()),
+                                out.numel(), quant_level, stream);
+    } else {
+      fa->allreduce<quickreduce::nv_bfloat16, false>(
+          reinterpret_cast<quickreduce::nv_bfloat16*>(inp.data_ptr()),
+          reinterpret_cast<quickreduce::nv_bfloat16*>(out.data_ptr()),
+          out.numel(), quant_level, stream);
+    }
+  } else {
+    throw std::runtime_error(
+        "quick allreduce only supports float16 and bfloat16");
+  }
+}
+
+int64_t qr_max_size() {
+  // The default is 2GB (2,147,483,648 bytes)
+  return static_cast<int64_t>(std::numeric_limits<int32_t>::max()) + 1;
+}
+
+  #define INSTANTIATE_FOR_WORLDSIZE(T, Codec, cast_bf2half)       \
+    template struct quickreduce::AllReduceTwoshot<T, Codec<T, 2>, \
+                                                  cast_bf2half>;  \
+    template struct quickreduce::AllReduceTwoshot<T, Codec<T, 4>, \
+                                                  cast_bf2half>;  \
+    template struct quickreduce::AllReduceTwoshot<T, Codec<T, 8>, cast_bf2half>;
+
+INSTANTIATE_FOR_WORLDSIZE(quickreduce::nv_bfloat16, quickreduce::CodecFP, false)
+INSTANTIATE_FOR_WORLDSIZE(quickreduce::nv_bfloat16, quickreduce::CodecQ4, false)
+INSTANTIATE_FOR_WORLDSIZE(quickreduce::nv_bfloat16, quickreduce::CodecQ6, false)
+INSTANTIATE_FOR_WORLDSIZE(quickreduce::nv_bfloat16, quickreduce::CodecQ8, false)
+INSTANTIATE_FOR_WORLDSIZE(quickreduce::nv_bfloat16, quickreduce::CodecFP, true)
+INSTANTIATE_FOR_WORLDSIZE(quickreduce::nv_bfloat16, quickreduce::CodecQ4, true)
+INSTANTIATE_FOR_WORLDSIZE(quickreduce::nv_bfloat16, quickreduce::CodecQ6, true)
+INSTANTIATE_FOR_WORLDSIZE(quickreduce::nv_bfloat16, quickreduce::CodecQ8, true)
+
+INSTANTIATE_FOR_WORLDSIZE(half, quickreduce::CodecFP, false)
+INSTANTIATE_FOR_WORLDSIZE(half, quickreduce::CodecQ4, false)
+INSTANTIATE_FOR_WORLDSIZE(half, quickreduce::CodecQ6, false)
+INSTANTIATE_FOR_WORLDSIZE(half, quickreduce::CodecQ8, false)
+
+#endif  // USE_ROCM

csrc/ops.h

@@ -360,3 +360,14 @@ std::tuple<int64_t, torch::Tensor> allocate_shared_buffer_and_handle(
     int64_t size);
 int64_t open_mem_handle(torch::Tensor& mem_handle);
 void free_shared_buffer(int64_t buffer);
+
+#ifdef USE_ROCM
+fptr_t init_custom_qr(int64_t rank, int64_t world_size,
+                      std::optional<int64_t> qr_max_size = std::nullopt);
+void qr_destroy(fptr_t _fa);
+torch::Tensor qr_get_handle(fptr_t _fa);
+void qr_open_handles(fptr_t _fa, const std::vector<torch::Tensor>& handles);
+void qr_all_reduce(fptr_t _fa, torch::Tensor& inp, torch::Tensor& out,
+                   int64_t quant_level, bool cast_bf2half = false);
+int64_t qr_max_size();
+#endif

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8_dispatch.cuh

@@ -29,40 +29,40 @@ struct sm100_fp8_config_default {
 template <typename InType, typename OutType,
           template <typename, typename, typename> typename Epilogue>
 struct sm100_fp8_config_M256 {
-  // M in (128, 256]
+  // M in (64, 256]
   static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
   using KernelSchedule = cutlass::gemm::collective::KernelScheduleAuto;
   using EpilogueSchedule = cutlass::epilogue::collective::EpilogueScheduleAuto;
   using TileShape = Shape<_128, _128, _128>;
-  using ClusterShape = Shape<_2, _2, _1>;
+  using ClusterShape = Shape<_2, _1, _1>;
   using Cutlass3xGemm =
       cutlass_3x_gemm_sm100<InType, OutType, Epilogue, TileShape, ClusterShape,
                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
           template <typename, typename, typename> typename Epilogue>
-struct sm100_fp8_config_M128 {
-  // M in (64, 128]
+struct sm100_fp8_config_M64 {
+  // M in (16, 64]
   static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
   using KernelSchedule = cutlass::gemm::collective::KernelScheduleAuto;
   using EpilogueSchedule = cutlass::epilogue::collective::EpilogueScheduleAuto;
-  using TileShape = Shape<_128, _128, _256>;
-  using ClusterShape = Shape<_2, _4, _1>;
+  using TileShape = Shape<_64, _64, _128>;
+  using ClusterShape = Shape<_1, _1, _1>;
   using Cutlass3xGemm =
       cutlass_3x_gemm_sm100<InType, OutType, Epilogue, TileShape, ClusterShape,
                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
           template <typename, typename, typename> typename Epilogue>
-struct sm100_fp8_config_M64 {
-  // M in [1, 64]
+struct sm100_fp8_config_M16 {
+  // M in [1, 16]
   static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
   using KernelSchedule = cutlass::gemm::collective::KernelScheduleAuto;
   using EpilogueSchedule = cutlass::epilogue::collective::EpilogueScheduleAuto;
-  using TileShape = Shape<_64, _64, _256>;
-  using ClusterShape = Shape<_1, _8, _1>;
+  using TileShape = Shape<_64, _64, _128>;
+  using ClusterShape = Shape<_1, _4, _1>;
   using Cutlass3xGemm =
       cutlass_3x_gemm_sm100<InType, OutType, Epilogue, TileShape, ClusterShape,
                             KernelSchedule, EpilogueSchedule>;
@@ -82,27 +82,27 @@ inline void cutlass_gemm_sm100_fp8_dispatch(torch::Tensor& out,
   using Cutlass3xGemmDefault =
       typename sm100_fp8_config_default<InType, OutType,
                                         Epilogue>::Cutlass3xGemm;
+  using Cutlass3xGemmM16 =
+      typename sm100_fp8_config_M16<InType, OutType, Epilogue>::Cutlass3xGemm;
   using Cutlass3xGemmM64 =
       typename sm100_fp8_config_M64<InType, OutType, Epilogue>::Cutlass3xGemm;
-  using Cutlass3xGemmM128 =
-      typename sm100_fp8_config_M128<InType, OutType, Epilogue>::Cutlass3xGemm;
   using Cutlass3xGemmM256 =
       typename sm100_fp8_config_M256<InType, OutType, Epilogue>::Cutlass3xGemm;
 
   uint32_t const m = a.size(0);
   uint32_t const mp2 =
-      std::max(static_cast<uint32_t>(64), next_pow_2(m));  // next power of 2
+      std::max(static_cast<uint32_t>(16), next_pow_2(m));  // next power of 2
 
-  if (mp2 <= 64) {
-    // m in [1, 64]
-    return cutlass_gemm_caller<Cutlass3xGemmM64>(
+  if (mp2 <= 16) {
+    // m in [1, 16]
+    return cutlass_gemm_caller<Cutlass3xGemmM16>(
         out, a, b, std::forward<EpilogueArgs>(args)...);
-  } else if (mp2 <= 128) {
-    // m in (64, 128]
-    return cutlass_gemm_caller<Cutlass3xGemmM128>(
+  } else if (mp2 <= 64) {
+    // m in (16, 64]
+    return cutlass_gemm_caller<Cutlass3xGemmM64>(
         out, a, b, std::forward<EpilogueArgs>(args)...);
   } else if (mp2 <= 256) {
-    // m in (128, 256]
+    // m in (64, 256]
     return cutlass_gemm_caller<Cutlass3xGemmM256>(
         out, a, b, std::forward<EpilogueArgs>(args)...);
   } else {

csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu

@@ -241,7 +241,7 @@ void get_cutlass_moe_mm_data(
   // mm to run it for.
   int32_t version_num = get_sm_version_num();
 #if (defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90) || \
-    (defined ENABLE_SCALED_MM_SM100 && ENABLE_SCALED_MM_SM90)
+    (defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100)
   get_cutlass_moe_mm_data_caller(topk_ids, expert_offsets, problem_sizes1,
                                  problem_sizes2, input_permutation,
                                  output_permutation, num_experts, n, k,
@@ -252,7 +252,7 @@ void get_cutlass_moe_mm_data(
       false,
       "No compiled get_cutlass_moe_mm_data: no cutlass_scaled_mm kernel for "
       "CUDA device capability: ",
-      version_num, ". Required capability: 90");
+      version_num, ". Required capability: 90 or 100");
 }
 
 void get_cutlass_pplx_moe_mm_data(torch::Tensor& expert_offsets,
@@ -265,7 +265,8 @@ void get_cutlass_pplx_moe_mm_data(torch::Tensor& expert_offsets,
   // This function currently gets compiled only if we have a valid cutlass moe
   // mm to run it for.
   int32_t version_num = get_sm_version_num();
-#if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90
+#if (defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90) || \
+    (defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100)
   get_cutlass_pplx_moe_mm_data_caller(expert_offsets, problem_sizes1,
                                       problem_sizes2, expert_num_tokens,
                                       num_local_experts, padded_m, n, k);
@@ -275,7 +276,7 @@ void get_cutlass_pplx_moe_mm_data(torch::Tensor& expert_offsets,
       false,
       "No compiled get_cutlass_pplx_moe_mm_data: no cutlass_scaled_mm kernel "
       "for CUDA device capability: ",
-      version_num, ". Required capability: 90");
+      version_num, ". Required capability: 90 or 100");
 }
 
 void cutlass_scaled_mm_azp(torch::Tensor& c, torch::Tensor const& a,