Merge branch 'vllm-project:main' into serialize-multimodal-kwargs

Author: p88h

.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/

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"

benchmarks/benchmark_throughput.py

@@ -597,18 +597,26 @@ def validate_args(args):
     parser.add_argument(
         "--prefix-len",
         type=int,
-        default=0,
-        help="Number of fixed prefix tokens before the random "
-        "context in a request (default: 0).",
-    )
+        default=None,
+        help=f"Number of prefix tokens to be used in RandomDataset "
+        "and SonnetDataset. For RandomDataset, the total input "
+        "length is the sum of prefix-len (default: "
+        f"{RandomDataset.DEFAULT_PREFIX_LEN}) and a random context length "
+        "sampled from [input_len * (1 - range_ratio), "
+        "input_len * (1 + range_ratio)]. For SonnetDataset, "
+        f"prefix_len (default: {SonnetDataset.DEFAULT_PREFIX_LEN}) "
+        "controls how much of the input is fixed lines versus "
+        "random lines, but the total input length remains approximately "
+        "input_len tokens.")
     # random dataset
     parser.add_argument(
         "--random-range-ratio",
         type=float,
-        default=0.0,
-        help="Range ratio for sampling input/output length, "
-        "used only for RandomDataset. Must be in the range [0, 1) to define "
-        "a symmetric sampling range "
+        default=None,
+        help=f"Range ratio (default : {RandomDataset.DEFAULT_RANGE_RATIO}) "
+        "for sampling input/output length, "
+        "used only for RandomDataset. Must be in the range [0, 1) to "
+        "define a symmetric sampling range "
         "[length * (1 - range_ratio), length * (1 + range_ratio)].",
     )
 

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);
+
+}