Support W4A8 quantization for vllm

.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

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"

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,

csrc/quantization/marlin/dense/common/base.h

@@ -0,0 +1,32 @@
+/*
+ * Modified by HandH1998
+ * Modified by Neural Magic
+ * Copyright (C) Marlin.2024 Elias Frantar
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *         http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+constexpr int ceildiv(int a, int b) { return (a + b - 1) / b; }
+
+// Instances of `Vec` are used to organize groups of >>registers<<, as needed
+// for instance as inputs to tensor core operations. Consequently, all
+// corresponding index accesses must be compile-time constants, which is why we
+// extensively use `#pragma unroll` throughout the kernel code to guarantee
+// this.
+template <typename T, int n>
+struct Vec {
+  T elems[n];
+  __device__ T& operator[](int i) { return elems[i]; }
+};

csrc/quantization/marlin/dense/common/mem.h

@@ -0,0 +1,89 @@
+/*
+ * Modified by HandH1998
+ * Modified by Neural Magic
+ * Copyright (C) Marlin.2024 Elias Frantar
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *         http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+// Predicated asynchronous global->shared copy; used for inputs A where we apply
+// predication to handle batchsizes that are not multiples of 16.
+__device__ inline void cp_async4_pred(void* smem_ptr, const void* glob_ptr,
+                                      bool pred = true) {
+  const int BYTES = 16;
+  uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
+  asm volatile(
+      "{\n"
+      "   .reg .pred p;\n"
+      "   setp.ne.b32 p, %0, 0;\n"
+      "   @p cp.async.cg.shared.global [%1], [%2], %3;\n"
+      "}\n" ::"r"((int)pred),
+      "r"(smem), "l"(glob_ptr), "n"(BYTES));
+}
+
+// Asynchronous global->shared copy
+__device__ inline void cp_async4(void* smem_ptr, const void* glob_ptr) {
+  const int BYTES = 16;
+  uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
+  asm volatile(
+      "{\n"
+      "   cp.async.cg.shared.global [%0], [%1], %2;\n"
+      "}\n" ::"r"(smem),
+      "l"(glob_ptr), "n"(BYTES));
+}
+
+// Async copy fence.
+__device__ inline void cp_async_fence() {
+  asm volatile("cp.async.commit_group;\n" ::);
+}
+
+// Wait until at most `n` async copy stages are still pending.
+template <int n>
+__device__ inline void cp_async_wait() {
+  asm volatile("cp.async.wait_group %0;\n" ::"n"(n));
+}
+
+// Wait until barrier reaches `count`, then lock for current threadblock.
+__device__ inline void barrier_acquire(int* lock, int count) {
+  if (threadIdx.x == 0) {
+    int state = -1;
+    do
+      // Guarantee that subsequent writes by this threadblock will be visible
+      // globally.
+      asm volatile("ld.global.acquire.gpu.b32 %0, [%1];\n"
+                   : "=r"(state)
+                   : "l"(lock));
+    while (state != count);
+  }
+  __syncthreads();
+}
+
+// Release barrier and increment visitation count.
+__device__ inline void barrier_release(int* lock, bool reset = false) {
+  __syncthreads();
+  if (threadIdx.x == 0) {
+    if (reset) {
+      lock[0] = 0;
+      return;
+    }
+    int val = 1;
+    // Make sure that all writes since acquiring this barrier are visible
+    // globally, while releasing the barrier.
+    asm volatile("fence.acq_rel.gpu;\n");
+    asm volatile("red.relaxed.gpu.global.add.s32 [%0], %1;\n"
+                 :
+                 : "l"(lock), "r"(val));
+  }
+}

csrc/quantization/marlin/dense/marlin_cuda_kernel.cu

@@ -25,30 +25,22 @@
 
 #include <iostream>
 
+#include "common/base.h"
+
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+  #include "common/mem.h"
+#endif
+
 template <typename T>
 inline std::string str(T x) {
   return std::to_string(x);
 }
 
 namespace marlin_dense {
 
-constexpr int ceildiv(int a, int b) { return (a + b - 1) / b; }
-
 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
 
-// Instances of `Vec` are used to organize groups of >>registers<<, as needed
-// for instance as inputs to tensor core operations. Consequently, all
-// corresponding index accesses must be compile-time constants, which is why we
-// extensively use `#pragma unroll` throughout the kernel code to guarantee
-// this.
-template <typename T, int n>
-struct Vec {
-  T elems[n];
-  __device__ T& operator[](int i) { return elems[i]; }
-};
-
 using I4 = Vec<int, 4>;
-
 // Matrix fragments for tensor core instructions; their precise layout is
 // documented here:
 // https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#matrix-fragments-for-mma-m16n8k16-with-floating-point-type
@@ -57,43 +49,6 @@ using FragB = Vec<half2, 2>;
 using FragC = Vec<float, 4>;
 using FragS = Vec<half2, 1>;  // quantization scales
 
-// Predicated asynchronous global->shared copy; used for inputs A where we apply
-// predication to handle batchsizes that are not multiples of 16.
-__device__ inline void cp_async4_pred(void* smem_ptr, const void* glob_ptr,
-                                      bool pred = true) {
-  const int BYTES = 16;
-  uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
-  asm volatile(
-      "{\n"
-      "   .reg .pred p;\n"
-      "   setp.ne.b32 p, %0, 0;\n"
-      "   @p cp.async.cg.shared.global [%1], [%2], %3;\n"
-      "}\n" ::"r"((int)pred),
-      "r"(smem), "l"(glob_ptr), "n"(BYTES));
-}
-
-// Asynchronous global->shared copy
-__device__ inline void cp_async4(void* smem_ptr, const void* glob_ptr) {
-  const int BYTES = 16;
-  uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
-  asm volatile(
-      "{\n"
-      "   cp.async.cg.shared.global [%0], [%1], %2;\n"
-      "}\n" ::"r"(smem),
-      "l"(glob_ptr), "n"(BYTES));
-}
-
-// Async copy fence.
-__device__ inline void cp_async_fence() {
-  asm volatile("cp.async.commit_group;\n" ::);
-}
-
-// Wait until at most `n` async copy stages are still pending.
-template <int n>
-__device__ inline void cp_async_wait() {
-  asm volatile("cp.async.wait_group %0;\n" ::"n"(n));
-}
-
 // m16n8k16 tensor core mma instruction with fp16 inputs and fp32
 // output/accumulation.
 __device__ inline void mma(const FragA& a_frag, const FragB& frag_b,
@@ -164,39 +119,6 @@ __device__ inline void scale(FragB& frag_b, FragS& frag_s, int i) {
   frag_b[1] = __hmul2(frag_b[1], s);
 }
 
-// Wait until barrier reaches `count`, then lock for current threadblock.
-__device__ inline void barrier_acquire(int* lock, int count) {
-  if (threadIdx.x == 0) {
-    int state = -1;
-    do
-      // Guarantee that subsequent writes by this threadblock will be visible
-      // globally.
-      asm volatile("ld.global.acquire.gpu.b32 %0, [%1];\n"
-                   : "=r"(state)
-                   : "l"(lock));
-    while (state != count);
-  }
-  __syncthreads();
-}
-
-// Release barrier and increment visitation count.
-__device__ inline void barrier_release(int* lock, bool reset = false) {
-  __syncthreads();
-  if (threadIdx.x == 0) {
-    if (reset) {
-      lock[0] = 0;
-      return;
-    }
-    int val = 1;
-    // Make sure that all writes since acquiring this barrier are visible
-    // globally, while releasing the barrier.
-    asm volatile("fence.acq_rel.gpu;\n");
-    asm volatile("red.relaxed.gpu.global.add.s32 [%0], %1;\n"
-                 :
-                 : "l"(lock), "r"(val));
-  }
-}
-
 template <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