merge: branch 'main' of github.com:vllm-project/vllm into feat/support-thinking-struct-outputs

Author: aarnphm

.buildkite/scripts/hardware_ci/run-cpu-test-ppc64le.sh

@@ -5,10 +5,34 @@
 set -ex
 
 # Setup cleanup
-remove_docker_container() { docker rm -f cpu-test || true; docker system prune -f; }
+remove_docker_container() { podman rm -f cpu-test-ubi9-ppc || true; podman system prune -f; }
 trap remove_docker_container EXIT
 remove_docker_container
 
 # Try building the docker image
-docker build -t cpu-test -f docker/Dockerfile.ppc64le .
+podman build -t cpu-test-ubi9-ppc -f docker/Dockerfile.ppc64le .
+
+# Run the image
+podman run -itd --entrypoint /bin/bash -v /tmp/:/root/.cache/huggingface --privileged=true --network host -e HF_TOKEN --name cpu-test-ubi9-ppc cpu-test-ubi9-ppc
+
+function cpu_tests() {
+
+  # offline inference
+  podman exec cpu-test-ubi9-ppc bash -c "
+    set -e
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
+
+  # Run basic model test
+  podman exec cpu-test-ubi9-ppc bash -c "
+    set -e
+    pip install pytest pytest-asyncio einops peft Pillow soundfile transformers_stream_generator matplotlib
+    pip install sentence-transformers datamodel_code_generator
+    pytest -v -s tests/models/embedding/language/test_cls_models.py::test_classification_models[float-jason9693/Qwen2.5-1.5B-apeach]
+    pytest -v -s tests/models/embedding/language/test_embedding.py::test_models[half-BAAI/bge-base-en-v1.5]
+    pytest -v -s tests/models/encoder_decoder/language -m cpu_model"
+}
+
+# All of CPU tests are expected to be finished less than 40 mins.
+export -f cpu_tests
+timeout 40m bash -c cpu_tests
 

.buildkite/scripts/hardware_ci/run-cpu-test-s390x.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+
+# This script build the CPU docker image and run the offline inference inside the container.
+# It serves a sanity check for compilation and basic model usage.
+set -ex
+
+# Setup cleanup
+remove_docker_container() { docker rm -f cpu-test || true; docker system prune -f; }
+trap remove_docker_container EXIT
+remove_docker_container
+
+# Try building the docker image
+docker build -t cpu-test -f docker/Dockerfile.s390x .

.buildkite/test-pipeline.yaml

@@ -341,6 +341,13 @@ steps:
   commands:
   - bash scripts/run-benchmarks.sh
 
+- label: Benchmarks CLI Test # 10min
+  source_file_dependencies:
+  - vllm/
+  - tests/benchmarks/
+  commands:
+  - pytest -v -s benchmarks/
+
 - label: Quantization Test # 33min
   source_file_dependencies:
   - csrc/
@@ -393,8 +400,9 @@ steps:
     - pytest -v -s models/test_transformers.py
     - pytest -v -s models/test_registry.py
     # V1 Test: https://github.com/vllm-project/vllm/issues/14531
-    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4'
+    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2'
     - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'llama4'
+    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'plamo2'
 
 - label: Language Models Test (Standard) # 32min
   #mirror_hardwares: [amd]
@@ -404,6 +412,8 @@ steps:
   - tests/models/embedding/language
   - tests/models/encoder_decoder/language
   commands:
+    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
+    - pip install causal-conv1d
     - pytest -v -s models/decoder_only/language -m 'core_model or quant_model'
     - pytest -v -s models/embedding/language -m core_model
 
@@ -415,6 +425,8 @@ steps:
   - tests/models/embedding/language
   - tests/models/encoder_decoder/language
   commands:
+    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
+    - pip install causal-conv1d
     - pytest -v -s models/decoder_only/language -m 'not core_model and not quant_model'
     - pytest -v -s models/embedding/language -m 'not core_model'
 
@@ -540,6 +552,7 @@ steps:
   # - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
   - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
   - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/test_disagg.py
+  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
 
 - label: Plugin Tests (2 GPUs) # 40min
   working_dir: "/vllm-workspace/tests"

.gitignore

@@ -203,3 +203,6 @@ benchmarks/**/*.json
 # Linting
 actionlint
 shellcheck*/
+
+# Ingore moe/marlin_moe gen code
+csrc/moe/marlin_moe_wna16/kernel_*

.pre-commit-config.yaml

@@ -11,7 +11,6 @@ repos:
   hooks:
   - id: yapf
     args: [--in-place, --verbose]
-    additional_dependencies: [toml] # TODO: Remove when yapf is upgraded
 - repo: https://github.com/astral-sh/ruff-pre-commit
   rev: v0.9.3
   hooks:

CMakeLists.txt

@@ -609,21 +609,51 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
   cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
   if (MARLIN_MOE_ARCHS)
-    set(MARLIN_MOE_SRC
-        "csrc/moe/marlin_kernels/marlin_moe_kernel.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4b8.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4b8.cu"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku8b128.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku8b128.cu"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4.cu"
-        "csrc/moe/marlin_moe_ops.cu")
 
+    #
+    # For the Marlin MOE kernels we automatically generate sources for various
+    # preselected input type pairs and schedules.
+    # Generate sources:
+    set(MOE_MARLIN_GEN_SCRIPT
+      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/moe/marlin_moe_wna16/generate_kernels.py)
+    file(MD5 ${MOE_MARLIN_GEN_SCRIPT} MOE_MARLIN_GEN_SCRIPT_HASH)
+
+    message(STATUS "Marlin MOE generation script hash: ${MOE_MARLIN_GEN_SCRIPT_HASH}")
+    message(STATUS "Last run Marlin MOE generate script hash: $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH}")
+
+    if (NOT DEFINED CACHE{MOE_MARLIN_GEN_SCRIPT_HASH}
+        OR NOT $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MOE_MARLIN_GEN_SCRIPT_HASH})
+      execute_process(
+        COMMAND ${CMAKE_COMMAND} -E env
+        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$PYTHONPATH
+          ${Python_EXECUTABLE} ${MOE_MARLIN_GEN_SCRIPT}
+        RESULT_VARIABLE moe_marlin_generation_result
+        OUTPUT_VARIABLE moe_marlin_generation_output
+        OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log
+        ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log
+      )
+
+      if (NOT moe_marlin_generation_result EQUAL 0)
+        message(FATAL_ERROR "Marlin MOE generation failed."
+                            " Result: \"${moe_marlin_generation_result}\""
+                            "\nCheck the log for details: "
+                            "${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log")
+      else()
+        set(MOE_MARLIN_GEN_SCRIPT_HASH ${MOE_MARLIN_GEN_SCRIPT_HASH}
+            CACHE STRING "Last run Marlin MOE generate script hash" FORCE)
+        message(STATUS "Marlin MOE generation completed successfully.")
+      endif()
+    else()
+      message(STATUS "Marlin MOE generation script has not changed, skipping generation.")
+    endif()
+
+    file(GLOB MOE_WNAA16_MARLIN_SRC "csrc/moe/marlin_moe_wna16/*.cu")
     set_gencode_flags_for_srcs(
-      SRCS "${MARLIN_MOE_SRC}"
+      SRCS "${MOE_WNAA16_MARLIN_SRC}"
       CUDA_ARCHS "${MARLIN_MOE_ARCHS}")
 
-    list(APPEND VLLM_MOE_EXT_SRC "${MARLIN_MOE_SRC}")
+    list(APPEND VLLM_MOE_EXT_SRC ${MOE_WNAA16_MARLIN_SRC})
+
     message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_ARCHS}")
   else()
     message(STATUS "Not building Marlin MOE kernels as no compatible archs found"

csrc/attention/merge_attn_states.cu

@@ -107,13 +107,14 @@ __global__ void merge_attn_states_kernel(
 
 #define LAUNCH_MERGE_ATTN_STATES(scalar_t, NUM_THREADS)                     \
   {                                                                         \
-    vllm::merge_attn_states_kernel<scalar_t, NUM_THREADS><<<grid, block>>>( \
-        reinterpret_cast<scalar_t*>(output.data_ptr()), output_lse_ptr,     \
-        reinterpret_cast<scalar_t*>(prefix_output.data_ptr()),              \
-        reinterpret_cast<float*>(prefix_lse.data_ptr()),                    \
-        reinterpret_cast<scalar_t*>(suffix_output.data_ptr()),              \
-        reinterpret_cast<float*>(suffix_lse.data_ptr()), num_tokens,        \
-        num_heads, head_size);                                              \
+    vllm::merge_attn_states_kernel<scalar_t, NUM_THREADS>                   \
+        <<<grid, block, 0, stream>>>(                                       \
+            reinterpret_cast<scalar_t*>(output.data_ptr()), output_lse_ptr, \
+            reinterpret_cast<scalar_t*>(prefix_output.data_ptr()),          \
+            reinterpret_cast<float*>(prefix_lse.data_ptr()),                \
+            reinterpret_cast<scalar_t*>(suffix_output.data_ptr()),          \
+            reinterpret_cast<float*>(suffix_lse.data_ptr()), num_tokens,    \
+            num_heads, head_size);                                          \
   }
 
 /*@brief Merges the attention states from prefix and suffix
@@ -122,10 +123,10 @@ __global__ void merge_attn_states_kernel(
  * @param output [n,h,d] The output tensor to store the merged attention states.
  * @param output_lse [h,d] Optional tensor to store the log-sum-exp values.
  * @param prefix_output [n,h,d] The prefix attention states.
- * @param prefix_lse [h,d] The log-sum-exp values for the prefix attention
+ * @param prefix_lse [h,n] The log-sum-exp values for the prefix attention
  * states.
  * @param suffix_output [n,h,d] The suffix attention states.
- * @param suffix_lse [h,d] The log-sum-exp values for the suffix attention
+ * @param suffix_lse [h,n] The log-sum-exp values for the suffix attention
  * states.
  */
 template <typename scalar_t>
@@ -146,13 +147,17 @@ void merge_attn_states_launcher(torch::Tensor& output,
   if (output_lse.has_value()) {
     output_lse_ptr = output_lse.value().data_ptr<float>();
   }
-  // process one pack elements per thread. float -> 4, half/bf16 -> 8
+  // Process one pack elements per thread. for float, the
+  // pack_size is 4 for half/bf16, the pack_size is 8.
   const uint threads_per_head = head_size / pack_size;
   const uint total_threads = num_tokens * num_heads * threads_per_head;
 
   dim3 block(NUM_THREADS);
   dim3 grid((total_threads + NUM_THREADS - 1) / NUM_THREADS);
 
+  const c10::cuda::OptionalCUDAGuard device_guard(prefix_output.device());
+  auto stream = at::cuda::getCurrentCUDAStream();
+
   LAUNCH_MERGE_ATTN_STATES(scalar_t, NUM_THREADS);
 }
 

csrc/moe/marlin_moe_wna16/generate_kernels.py

@@ -0,0 +1,103 @@
+# SPDX-License-Identifier: Apache-2.0
+import glob
+import itertools
+import os
+import subprocess
+
+import jinja2
+
+FILE_HEAD = """
+// auto generated by generate.py
+// clang-format off
+
+#include "kernel.h"
+#include "marlin_template.h"
+
+namespace MARLIN_NAMESPACE_NAME {
+""".strip()
+
+TEMPLATE = ("template __global__ void Marlin<"
+            "{{scalar_t}}, "
+            "{{w_type_id}}, "
+            "{{threads}}, "
+            "{{thread_m_blocks}}, "
+            "{{thread_n_blocks}}, "
+            "{{thread_k_blocks}}, "
+            "{{'true' if m_block_size_8 else 'false'}}, "
+            "{{stages}}, "
+            "{{'true' if has_act_order else 'false'}}, "
+            "{{'true' if has_zp else 'false'}}, "
+            "{{group_blocks}}, "
+            "{{'true' if is_zp_float else 'false'}}>"
+            "( MARLIN_KERNEL_PARAMS );")
+
+# int8 with zero point case (vllm::kU8) is also supported,
+# we don't add it to reduce wheel size.
+SCALAR_TYPES = ["vllm::kU4", "vllm::kU4B8", "vllm::kU8B128"]
+THREAD_CONFIGS = [(128, 128, 256), (64, 256, 256), (64, 128, 128)]
+
+THREAD_M_BLOCKS = [0.5, 1, 2, 3, 4]
+# group_blocks:
+#   = 0 : act order case
+#   = -1 : channelwise quantization
+#   > 0 : group_size=16*group_blocks
+GROUP_BLOCKS = [0, -1, 2, 4, 8]
+DTYPES = ["fp16", "bf16"]
+
+
+def remove_old_kernels():
+    for filename in glob.glob(os.path.dirname(__file__) + "/kernel_*.cu"):
+        subprocess.call(["rm", "-f", filename])
+
+
+def generate_new_kernels():
+    for scalar_type, dtype in itertools.product(SCALAR_TYPES, DTYPES):
+        has_zp = "B" not in scalar_type
+        all_template_str_list = []
+
+        for group_blocks, m_blocks, thread_configs in itertools.product(
+                GROUP_BLOCKS, THREAD_M_BLOCKS, THREAD_CONFIGS):
+
+            has_act_order = group_blocks == 0
+            if has_zp and has_act_order:
+                continue
+            if thread_configs[2] == 256:
+                if m_blocks <= 1 and thread_configs[0] != 128:
+                    continue
+                if m_blocks > 1 and thread_configs[0] != 64:
+                    continue
+
+            k_blocks = thread_configs[0] // 16
+            n_blocks = thread_configs[1] // 16
+            threads = thread_configs[2]
+
+            c_dtype = "half" if dtype == "fp16" else "nv_bfloat16"
+
+            template_str = jinja2.Template(TEMPLATE).render(
+                scalar_t=c_dtype,
+                w_type_id=scalar_type + ".id()",
+                threads=threads,
+                thread_m_blocks=max(m_blocks, 1),
+                thread_n_blocks=n_blocks,
+                thread_k_blocks=k_blocks,
+                m_block_size_8=m_blocks == 0.5,
+                stages="pipe_stages",
+                has_act_order=has_act_order,
+                has_zp=has_zp,
+                group_blocks=group_blocks,
+                is_zp_float=False,
+            )
+
+            all_template_str_list.append(template_str)
+
+        file_content = FILE_HEAD + "\n\n"
+        file_content += "\n\n".join(all_template_str_list) + "\n\n}\n"
+        filename = f"kernel_{dtype}_{scalar_type[6:].lower()}.cu"
+
+        with open(os.path.join(os.path.dirname(__file__), filename), "w") as f:
+            f.write(file_content)
+
+
+if __name__ == "__main__":
+    remove_old_kernels()
+    generate_new_kernels()

csrc/moe/marlin_moe_wna16/kernel.h

@@ -0,0 +1,44 @@
+
+#ifndef MARLIN_NAMESPACE_NAME
+  #define MARLIN_NAMESPACE_NAME marlin_moe_wna16
+#endif
+
+#include "quantization/gptq_marlin/marlin.cuh"
+#include "quantization/gptq_marlin/marlin_dtypes.cuh"
+#include "core/scalar_type.hpp"
+
+#define MARLIN_KERNEL_PARAMS                                                \
+  const int4 *__restrict__ A, const int4 *__restrict__ B,                   \
+      int4 *__restrict__ C, int4 *__restrict__ C_tmp,                       \
+      const int4 *__restrict__ scales_ptr, const int4 *__restrict__ zp_ptr, \
+      const int *__restrict__ g_idx,                                        \
+      const int32_t *__restrict__ sorted_token_ids_ptr,                     \
+      const int32_t *__restrict__ expert_ids_ptr,                           \
+      const int32_t *__restrict__ num_tokens_past_padded_ptr,               \
+      const float *__restrict__ topk_weights_ptr, int top_k,                \
+      bool mul_topk_weights, bool is_ep, int num_groups, int prob_m,        \
+      int prob_n, int prob_k, int *locks, bool use_atomic_add,              \
+      bool use_fp32_reduce
+
+namespace MARLIN_NAMESPACE_NAME {
+template <typename scalar_t,  // compute dtype, half or nv_float16
+          const vllm::ScalarTypeId w_type_id,  // weight ScalarType id
+          const int threads,          // number of threads in a threadblock
+          const int thread_m_blocks,  // number of 16x16 blocks in the m
+                                      // dimension (batchsize) of the
+                                      // threadblock
+          const int thread_n_blocks,  // same for n dimension (output)
+          const int thread_k_blocks,  // same for k dimension (reduction)
+          const bool m_block_size_8,  // whether m_block_size == 8
+                                      // only works when thread_m_blocks == 1
+          const int stages,  // number of stages for the async global->shared
+                             // fetch pipeline
+          const bool has_act_order,  // whether act_order is enabled
+          const bool has_zp,         // whether zero-points are enabled
+          const int group_blocks,    // number of consecutive 16x16 blocks
+                                     // with a separate quantization scale
+          const bool is_zp_float     // is zero point of float16 type?
+          >
+__global__ void Marlin(MARLIN_KERNEL_PARAMS);
+
+}