Merge branch 'vllm-project:main' into main

Author: haichuan1221

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-QQQ.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m HandH1998/QQQ-Llama-3-8b-g128 -b 32 -l 1000 -f 5 -t 1
+model_name: "HandH1998/QQQ-Llama-3-8b-g128"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.409
+  - name: "exact_match,flexible-extract"
+    value: 0.406
+limit: 1000
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/models-small.txt

@@ -7,3 +7,4 @@ Meta-Llama-3-8B-Instruct-Channelwise-compressed-tensors.yaml
 Minitron-4B-Base.yaml
 Qwen2-1.5B-Instruct-INT8-compressed-tensors.yaml
 Qwen2-1.5B-Instruct-FP8W8.yaml
+Meta-Llama-3-8B-QQQ.yaml

.github/workflows/scripts/build.sh

@@ -13,8 +13,6 @@ $python_executable -m pip install -r requirements-cuda.txt
 
 # Limit the number of parallel jobs to avoid OOM
 export MAX_JOBS=1
-# Make sure punica is built for the release (for LoRA)
-export VLLM_INSTALL_PUNICA_KERNELS=1
 # Make sure release wheels are built for the following architectures
 export TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 8.9 9.0+PTX"
 # Build

CMakeLists.txt

@@ -170,6 +170,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     "csrc/quantization/awq/gemm_kernels.cu"
     "csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
     "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
+    "csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu"
     "csrc/quantization/gptq_marlin/gptq_marlin.cu"
     "csrc/quantization/gptq_marlin/gptq_marlin_repack.cu"
     "csrc/quantization/gptq_marlin/awq_marlin_repack.cu"
@@ -222,61 +223,7 @@ define_gpu_extension_target(
   USE_SABI 3
   WITH_SOABI)
 
-#
-# _punica_C extension
-#
-
-set(VLLM_PUNICA_EXT_SRC
-  "csrc/punica/bgmv/bgmv_bf16_bf16_bf16.cu"
-  "csrc/punica/bgmv/bgmv_bf16_fp32_bf16.cu"
-  "csrc/punica/bgmv/bgmv_fp16_fp16_fp16.cu"
-  "csrc/punica/bgmv/bgmv_fp16_fp32_fp16.cu"
-  "csrc/punica/bgmv/bgmv_fp32_bf16_bf16.cu"
-  "csrc/punica/bgmv/bgmv_fp32_fp16_fp16.cu"
-  "csrc/punica/punica_ops.cu"
-  "csrc/punica/torch_bindings.cpp")
-
-#
-# Copy GPU compilation flags+update for punica
-#
-set(VLLM_PUNICA_GPU_FLAGS ${VLLM_GPU_FLAGS})
-list(REMOVE_ITEM VLLM_PUNICA_GPU_FLAGS
-  "-D__CUDA_NO_HALF_OPERATORS__"
-  "-D__CUDA_NO_HALF_CONVERSIONS__"
-  "-D__CUDA_NO_BFLOAT16_CONVERSIONS__"
-  "-D__CUDA_NO_HALF2_OPERATORS__")
-
-#
-# Filter out CUDA architectures < 8.0 for punica.
-#
-if (${VLLM_GPU_LANG} STREQUAL "CUDA")
-  set(VLLM_PUNICA_GPU_ARCHES)
-  foreach(ARCH ${VLLM_GPU_ARCHES})
-    string_to_ver(CODE_VER ${ARCH})
-    if (CODE_VER GREATER_EQUAL 8.0)
-      list(APPEND VLLM_PUNICA_GPU_ARCHES ${ARCH})
-    endif()
-  endforeach()
-  message(STATUS "Punica target arches: ${VLLM_PUNICA_GPU_ARCHES}")
-elseif(${VLLM_GPU_LANG} STREQUAL "HIP")
-  set(VLLM_PUNICA_GPU_ARCHES ${VLLM_GPU_ARCHES})
-  message(STATUS "Punica target arches: ${VLLM_PUNICA_GPU_ARCHES}")
-endif()
 
-if (VLLM_PUNICA_GPU_ARCHES)
-  define_gpu_extension_target(
-    _punica_C
-    DESTINATION vllm
-    LANGUAGE ${VLLM_GPU_LANG}
-    SOURCES ${VLLM_PUNICA_EXT_SRC}
-    COMPILE_FLAGS ${VLLM_PUNICA_GPU_FLAGS}
-    ARCHITECTURES ${VLLM_PUNICA_GPU_ARCHES}
-    USE_SABI 3
-    WITH_SOABI)
-else()
-  message(WARNING "Unable to create _punica_C target because none of the "
-    "requested architectures (${VLLM_GPU_ARCHES}) are supported, i.e. >= 8.0")
-endif()
 
 #
 # Add the `default` target which detects which extensions should be
@@ -300,12 +247,4 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
   message(STATUS "Enabling moe extension.")
   add_dependencies(default _moe_C)
 
-  # Enable punica if -DVLLM_INSTALL_PUNICA_KERNELS=ON or
-  # VLLM_INSTALL_PUNICA_KERNELS is set in the environment and
-  # there are supported target arches.
-  if (VLLM_PUNICA_GPU_ARCHES AND
-      (ENV{VLLM_INSTALL_PUNICA_KERNELS} OR VLLM_INSTALL_PUNICA_KERNELS))
-    message(STATUS "Enabling punica extension.")
-    add_dependencies(default _punica_C)
-  endif()
 endif()

Dockerfile

@@ -88,8 +88,6 @@ ENV MAX_JOBS=${max_jobs}
 # number of threads used by nvcc
 ARG nvcc_threads=8
 ENV NVCC_THREADS=$nvcc_threads
-# make sure punica kernels are built (for LoRA)
-ENV VLLM_INSTALL_PUNICA_KERNELS=1
 
 ARG buildkite_commit
 ENV BUILDKITE_COMMIT=${buildkite_commit}

Dockerfile.rocm

@@ -131,8 +131,7 @@ COPY . .
 RUN --mount=type=cache,target=/root/.cache/pip \
     python3 -m pip install --upgrade numba scipy huggingface-hub[cli]
 
-# Make sure punica kernels are built (for LoRA)
-ENV VLLM_INSTALL_PUNICA_KERNELS=1
+
 # Workaround for ray >= 2.10.0
 ENV RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES=1
 # Silences the HF Tokenizers warning

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

@@ -112,13 +112,20 @@ def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
     scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
 
     timers = []
-    # pytorch impl
+    # pytorch impl - bfloat16
     timers.append(
         bench_fn(a.to(dtype=torch.bfloat16, device="cuda"),
                  b.to(dtype=torch.bfloat16, device="cuda"), scale_a, scale_b,
                  torch.bfloat16, label, sub_label, pytorch_mm_impl,
                  "pytorch_bf16_bf16_bf16_matmul-no-scales"))
 
+    # pytorch impl - float16
+    timers.append(
+        bench_fn(a.to(dtype=torch.float16, device="cuda"),
+                 b.to(dtype=torch.float16, device="cuda"), scale_a, scale_b,
+                 torch.float16, label, sub_label, pytorch_mm_impl,
+                 "pytorch_fp16_fp16_fp16_matmul-no-scales"))
+
     # cutlass impl
     timers.append(
         bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,

csrc/ops.h

@@ -115,6 +115,13 @@ void cutlass_scaled_mm(torch::Tensor& out, torch::Tensor const& a,
                        torch::Tensor const& b_scales,
                        c10::optional<torch::Tensor> const& bias);
 
+torch::Tensor marlin_qqq_gemm(torch::Tensor const& a,
+                              torch::Tensor const& b_q_weight,
+                              torch::Tensor const& s_tok,
+                              torch::Tensor const& s_ch,
+                              torch::Tensor const& s_group,
+                              torch::Tensor& workspace, int64_t size_m,
+                              int64_t size_n, int64_t size_k);
 #endif
 
 void static_scaled_int8_quant(torch::Tensor& out, torch::Tensor const& input,