AQLM CUDA support (#3287)

Co-authored-by: mgoin <michael@neuralmagic.com>

CMakeLists.txt

@@ -173,6 +173,7 @@ set(VLLM_EXT_SRC
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_EXT_SRC
+    "csrc/quantization/aqlm/gemm_kernels.cu"
     "csrc/quantization/awq/gemm_kernels.cu"
     "csrc/quantization/marlin/marlin_cuda_kernel.cu"
     "csrc/custom_all_reduce.cu")

benchmarks/kernels/benchmark_aqlm.py

+import argparse
+import os
+import sys
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+
+from vllm._C import ops
+from vllm.model_executor.layers.quantization.aqlm import (
+    dequantize_weight, generic_dequantize_gemm, get_int_dtype,
+    optimized_dequantize_gemm)
+
+os.environ['CUDA_VISIBLE_DEVICES'] = '0'
+
+
+def torch_mult(
+        input: torch.Tensor,  #  [..., in_features]
+        weights: torch.Tensor,
+        scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+) -> torch.Tensor:
+    output = F.linear(input, weights)
+    return output
+
+
+def dequant_out_scale(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    output_partition_sizes: torch.IntTensor,
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+
+    weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
+
+    if bias is None:
+        output = F.linear(input, weights, bias)
+        orig_shape = output.shape
+        flattened_output = output.view(-1, output.size(-1))
+        f_scales = scales.view(-1, scales.shape[0])
+        b_scales = f_scales.expand(flattened_output.shape[0], -1)
+        flattened_output *= b_scales
+        return flattened_output.view(orig_shape)
+    else:
+        b_scales = scales.view(scales.shape[:-3] + (-1, )).expand(
+            -1, weights.shape[1])
+        weights *= b_scales
+        return F.linear(input, weights, bias)
+
+
+def dequant_weight_scale(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    output_partition_sizes: torch.IntTensor,
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+
+    weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
+
+    b_scales = scales.view(scales.shape[:-3] + (-1, )).expand(
+        -1, weights.shape[1])
+    weights *= b_scales
+    return F.linear(input, weights, bias)
+
+
+def dequant_no_scale(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    output_partition_sizes: torch.IntTensor,
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+
+    weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
+
+    return F.linear(input, weights, bias)
+
+
+# Compare the optimized 1x16 and 2x8 cuda decompression/dequant kernels against
+# the generic pytorch version.
+# Just visual comparison.
+def dequant_test(k: int, parts: torch.tensor, nbooks: int, bits: int) -> None:
+
+    n = parts.sum().item()
+
+    device = torch.device('cuda:0')
+
+    code_range = (1 << bits) // 2
+    ingroups = 8
+
+    codes = torch.randint(-code_range,
+                          code_range,
+                          size=(n, k // ingroups, nbooks),
+                          dtype=get_int_dtype(bits),
+                          device=device)
+
+    codebooks = torch.randn(size=(parts.shape[0] * nbooks, 1 << bits, 1, 8),
+                            dtype=torch.float16,
+                            device=device)
+
+    count = 0
+    for index in range(16):
+        for i in range(8):
+            for book in range(nbooks):
+                codebooks[book, index, 0, i] = count * (10**book)
+            count += 1
+
+    print("codes shape", codes.shape)
+
+    for i in range(16):
+        for book in range(nbooks):
+            codes[0, i, book] = i
+            codes[0, -i, book] = i
+
+    weights = dequantize_weight(codes, codebooks, None)
+    weights2 = ops.aqlm_dequant(codes, codebooks, parts)
+
+    print("weights shape:", weights.shape)
+    print("weights2 shape:", weights2.shape)
+
+    print("weights are:", weights)
+    print("weights2 are:", weights2)
+
+    print("first 128 weights are", weights[0, 0:128].to(torch.int32))
+    print("first 128 weights2 are:", weights2[0, 0:128].to(torch.int32))
+
+    print("last 128 weights are", weights[0, -128:])
+    print("last 128 weights2 are:", weights2[0, -128:])
+
+
+def main():
+
+    parser = argparse.ArgumentParser(description="Benchmark aqlm performance.")
+
+    # Add arguments
+    parser.add_argument("--nbooks",
+                        type=int,
+                        default=1,
+                        help="Number of codebooks (default: 1)")
+    parser.add_argument("--bits",
+                        type=int,
+                        default=16,
+                        help="Number of bits per code element (default: 16)")
+    parser.add_argument(
+        "--test",
+        type=bool,
+        default=False,
+        help="Run the decompression/dequant tester rather than benchmarking "
+        "(default: False)")
+
+    # Parse the arguments
+    args = parser.parse_args()
+
+    # Extract values
+    nbooks = args.nbooks
+    bits = args.bits
+
+    if args.test:
+        dequant_test(4096, torch.tensor((4096, )), nbooks, bits)
+        return
+
+    # Otherwise, benchmark.
+    methods = [
+        ops.aqlm_gemm,
+        dequant_out_scale,
+        generic_dequantize_gemm,
+        optimized_dequantize_gemm,
+        dequant_weight_scale,
+        torch_mult,
+        dequant_no_scale,
+    ]
+
+    filename = f"./aqlm_benchmark_{nbooks}x{bits}.csv"
+    print(f"writing benchmarks to file {filename}")
+    with open(filename, "w") as f:
+        sys.stdout = f
+
+        print('m | k | n | n parts', end='')
+        for method in methods:
+            print(f" | {method.__name__.replace('_', ' ')} (µs)", end='')
+        print('')
+
+        # These are reasonable prefill sizes.
+        ksandpartions = ((4096, (4096, 4096, 4096)), (4096, (4096, )),
+                         (4096, (11008, 11008)), (11008, (4096, )))
+
+        # reasonable ranges for m.
+        for m in [
+                1, 2, 4, 8, 10, 12, 14, 16, 24, 32, 48, 52, 56, 64, 96, 112,
+                128, 256, 512, 1024, 1536, 2048, 3072, 4096
+        ]:
+            print(f'{m}', file=sys.__stdout__)
+            for ksp in ksandpartions:
+                run_grid(m, ksp[0], torch.tensor(ksp[1]), nbooks, bits,
+                         methods)
+
+        sys.stdout = sys.__stdout__
+
+
+def run_grid(m: int, k: int, parts: torch.tensor, nbooks: int, bits: int,
+             methods):
+
+    # I didn't see visible improvements from increasing these, but feel free :)
+    num_warmup_trials = 1
+    num_trials = 1
+
+    num_calls = 100
+
+    # warmup.
+    for method in methods:
+        for _ in range(num_warmup_trials):
+            run_timing(
+                num_calls=num_calls,
+                m=m,
+                k=k,
+                parts=parts,
+                nbooks=nbooks,
+                bits=bits,
+                method=method,
+            )
+
+    n = parts.sum().item()
+    print(f'{m} | {k} | {n} | {parts.tolist()}', end='')
+
+    for method in methods:
+        best_time_us = 1e20
+        for _ in range(num_trials):
+            kernel_dur_ms = run_timing(
+                num_calls=num_calls,
+                m=m,
+                k=k,
+                parts=parts,
+                nbooks=nbooks,
+                bits=bits,
+                method=method,
+            )
+
+            kernel_dur_us = 1000 * kernel_dur_ms
+
+            if kernel_dur_us < best_time_us:
+                best_time_us = kernel_dur_us
+
+        print(f' | {kernel_dur_us:.0f}', end='')
+
+    print('')
+
+
+def run_timing(num_calls: int, m: int, k: int, parts: torch.tensor,
+               nbooks: int, bits: int, method) -> float:
+
+    n = parts.sum().item()
+
+    device = torch.device('cuda:0')
+
+    input = torch.randn((1, m, k), dtype=torch.float16, device=device)
+
+    code_range = (1 << bits) // 2
+    ingroups = 8
+
+    codes = torch.randint(-code_range,
+                          code_range,
+                          size=(n, k // ingroups, nbooks),
+                          dtype=get_int_dtype(bits),
+                          device=device)
+
+    codebooks = torch.randn(size=(parts.shape[0] * nbooks, 1 << bits, 1, 8),
+                            dtype=torch.float16,
+                            device=device)
+
+    scales = torch.randn(size=(n, 1, 1, 1), dtype=torch.float16, device=device)
+
+    # for comparison to just a pytorch mult.
+    weights = torch.randn((n, k), dtype=torch.float16, device=device)
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    start_event.record()
+
+    if method is torch_mult:
+        for i in range(num_calls):
+            torch_mult(input, weights, scales)
+    else:
+        for i in range(num_calls):
+            method(input, codes, codebooks, scales, parts, None)
+
+    end_event.record()
+    end_event.synchronize()
+
+    dur_ms = start_event.elapsed_time(end_event) / num_calls
+    return dur_ms
+
+
+if __name__ == "__main__":
+    sys.exit(main())

csrc/ops.h

@@ -86,6 +86,21 @@ void gelu_fast(
   torch::Tensor& input);
 
 #ifndef USE_ROCM
+torch::Tensor aqlm_gemm(
+  const torch::Tensor& input,
+  const torch::Tensor& codes,
+  const torch::Tensor& codebooks,
+  const torch::Tensor& scales,
+  const torch::Tensor& codebook_partition_sizes,
+  const std::optional<torch::Tensor>& bias
+);
+
+torch::Tensor aqlm_dequant(
+  const torch::Tensor& codes,
+  const torch::Tensor& codebooks,
+  const torch::Tensor& codebook_partition_sizes
+);
+
 torch::Tensor awq_gemm(
   torch::Tensor _in_feats,
   torch::Tensor _kernel,

csrc/pybind.cpp

@@ -63,6 +63,8 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
 
 // Quantization ops
 #ifndef USE_ROCM
+  ops.def("aqlm_gemm", &aqlm_gemm, "Quantized GEMM for AQLM");
+  ops.def("aqlm_dequant", &aqlm_dequant, "Decompression method for AQLM");
   ops.def("awq_gemm", &awq_gemm, "Quantized GEMM for AWQ");
   ops.def("marlin_gemm", &marlin_gemm, "Marlin Optimized Quantized GEMM for GPTQ");
   ops.def("awq_dequantize", &awq_dequantize, "Dequantization for AWQ");

csrc/quantization/aqlm/gemm_kernels.cu

+/*
+ * Modified by Neural Magic
+ * Adapted from https://github.com/Vahe1994/AQLM
+ *
+ * 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.
+ */
+
+#include <cuda.h>
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+#include <c10/cuda/CUDAStream.h>
+#include <c10/cuda/CUDAGuard.h>
+
+#include <iostream>
+#include <cstdlib>
+
+
+namespace vllm {
+namespace aqlm {
+
+__global__ void Code1x16MatVec(
+  const int4* __restrict__ A,
+  const int4* __restrict__ B,
+        int4* __restrict__ C,
+  const int4* __restrict__ codebook,
+  const int prob_m,
+  const int prob_k,
+  const int4 codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at most 3 long.
+  const int codebook_stride // as int4.
+) {
+  int a_gl_stride = prob_k / 8 / 8;
+  int a_gl_rd = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
+  bool pred = a_gl_rd < prob_m;
+
+  if (pred)
+  {
+    // advance to the correct codebook, this easy because we only multiply one column of the codebook.
+    auto codebook_size = &codebook_a_sizes.x;
+    while (a_gl_rd >= *codebook_size)
+    {
+        codebook += codebook_stride;
+        ++codebook_size;
+    }
+  }
+
+  int b_gl_rd = 0;
+  int c_gl_wr = a_gl_rd;
+  a_gl_rd = a_gl_stride * a_gl_rd + threadIdx.x % 32;
+  int a_gl_end = a_gl_rd + a_gl_stride - threadIdx.x % 32;
+
+  __shared__ int4 sh_b[32 * 9];
+  float res = 0;
+
+  int iters = (prob_k / 8 + 8 * 32 - 1) / (8 * 32);
+  while (iters--) {
+    // We pad shared memory to avoid bank conflicts during reads
+    __syncthreads();
+    for (int i = threadIdx.x; i < 32 * 8; i += blockDim.x) {
+      if (b_gl_rd + i < prob_k / 8)
+        sh_b[9 * (i / 8) + i % 8] = B[b_gl_rd + i];
+    }
+    __syncthreads();
+    b_gl_rd += 32 * 8;
+
+    int b_sh_rd = 9 * (threadIdx.x % 32);
+    if (pred && a_gl_rd < a_gl_end) {
+      const uint16_t* enc = reinterpret_cast<const uint16_t*>(&A[a_gl_rd]);
+      #pragma unroll
+      for (int i = 0; i < 8; i++) {
+        uint32_t dec[4];
+        // We bypass the L1 cache to avoid massive amounts of memory streaming that doesn't
+        // actually help us; this brings > 2x speedup.
+        asm volatile (
+          "ld.cg.global.v4.u32 {%0, %1, %2, %3}, [%4];"
+          : "=r"(dec[0]), "=r"(dec[1]), "=r"(dec[2]), "=r"(dec[3])
+          : "l"((void*) &codebook[enc[i]])
+        );
+        half2* a = reinterpret_cast<half2*>(&dec);
+        half2* b = reinterpret_cast<half2*>(&sh_b[b_sh_rd]);
+        half2 res2 = {};
+        #pragma unroll
+        for (int j = 0; j < 4; j++)
+          res2 = __hfma2(a[j], b[j], res2);
+        res += __half2float(res2.x) + __half2float(res2.y);
+        b_sh_rd++;
+      }
+      a_gl_rd += 32;
+    }
+  }
+
+  if (pred) {
+    #pragma unroll
+    for (int i = 16; i > 0; i /= 2)
+      res += __shfl_down_sync(0xffffffff, res, i);
+    if (threadIdx.x % 32 == 0)
+      reinterpret_cast<__half*>(C)[c_gl_wr] = __float2half(res);
+  }
+}
+
+__global__ void Code2x8MatVec(
+  const int4* __restrict__ A,
+  const int4* __restrict__ B,
+        int4* __restrict__ C,
+  const int4* __restrict__ codebook,
+  int prob_m,
+  int prob_k,
+  const int4 codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at most 3 long.
+  const int codebook_stride // as int4.
+
+) {
+  int a_gl_stride = prob_k / 8 / 8;
+  int a_gl_rd = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
+  bool pred = a_gl_rd < prob_m;
+
+  if (pred)
+  {
+    // advance to the correct codebook, this easy because we only multiply one column of the codebook.
+    auto codebook_size = &codebook_a_sizes.x;
+    while (a_gl_rd >= *codebook_size)
+    {
+        codebook += codebook_stride;
+        ++codebook_size;
+    }
+  }
+
+  int b_gl_rd = 0;
+  int c_gl_wr = a_gl_rd;
+  a_gl_rd = a_gl_stride * a_gl_rd + threadIdx.x % 32;
+  int a_gl_end = a_gl_rd + a_gl_stride - threadIdx.x % 32;
+  int lane = threadIdx.x % 8;
+
+  extern __shared__ int4 sh[];
+  int4* sh_b = sh;
+  int4* sh_code = sh_b + 32 * 9;
+  int4* sh_code0 = sh_code;
+  int4* sh_code1 = sh_code + 256 * 8;
+
+  for (int i = threadIdx.x; i < 2 * 256; i += blockDim.x) {
+    int4 dec = codebook[i];
+    #pragma unroll
+    for (int j = 0; j < 8; j++)
+      sh_code[8 * i + (j + lane) % 8] = dec;
+  }
+  __syncthreads();
+
+  float res = 0;
+
+  int iters = (prob_k / 8 + 8 * 32 - 1) / (8 * 32);
+  while (iters--) {
+    // We pad shared memory to avoid bank conflicts during reads
+    __syncthreads();
+    for (int i = threadIdx.x; i < 32 * 8; i += blockDim.x) {
+      if (b_gl_rd + i < prob_k / 8)
+        sh_b[9 * (i / 8) + i % 8] = B[b_gl_rd + i];
+    }
+    __syncthreads();
+    b_gl_rd += 32 * 8;
+
+    int b_sh_rd = 9 * (threadIdx.x % 32);
+    if (pred && a_gl_rd < a_gl_end) {
+      const uint8_t* enc = reinterpret_cast<const uint8_t*>(&A[a_gl_rd]);
+      #pragma unroll
+      for (int i = 0; i < 8; i++) {
+        half2* a0 = reinterpret_cast<half2*>(&sh_code0[8 * enc[2 * i + 0] + lane]);
+        half2* a1 = reinterpret_cast<half2*>(&sh_code1[8 * enc[2 * i + 1] + lane]);
+        half2*  b = reinterpret_cast<half2*>(&sh_b[b_sh_rd]);
+        half2 res2 = {};
+        #pragma unroll
+        for (int j = 0; j < 4; j++)
+          res2 = __hfma2(__hadd2(a0[j], a1[j]), b[j], res2);
+        res += __half2float(res2.x) + __half2float(res2.y);
+        b_sh_rd++;
+      }
+      a_gl_rd += 32;
+    }
+  }
+
+  if (pred) {
+    #pragma unroll
+    for (int i = 16; i > 0; i /= 2)
+      res += __shfl_down_sync(0xffffffff, res, i);
+    if (threadIdx.x % 32 == 0)
+      reinterpret_cast<__half*>(C)[c_gl_wr] = __float2half(res);
+  }
+}
+
+
+__global__ void Code1x16Dequant(
+  const int4* __restrict__ A,
+        int4* __restrict__ C,
+  const int4* __restrict__ codebook,
+  int prob_m,
+  int prob_k,
+  const int4 codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at most 3 long, sums to m.
+  const int codebook_stride // as int4
+) {
+  int a_gl_stride = prob_k / 8 / 8;
+  int a_gl_rd = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
+  bool pred = a_gl_rd < prob_m;
+
+  if (pred)
+  {
+    // advance to the correct codebook, this easy because we only multiply one column of the codebook.
+    auto codebook_size = &codebook_a_sizes.x;
+    while (a_gl_rd >= *codebook_size)
+    {
+        codebook += codebook_stride;
+        ++codebook_size;
+    }
+  }
+
+  a_gl_rd = a_gl_stride * a_gl_rd + threadIdx.x % 32;
+  int a_gl_end = a_gl_rd + a_gl_stride - threadIdx.x % 32;
+
+  int c_gl_stride = prob_k / 8;
+  int c_gl_wr = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
+  c_gl_wr = c_gl_stride * c_gl_wr + (threadIdx.x % 32) * 8;
+
+  int iters = (prob_k / 8 - 1) / (8 * 32) + 1;
+  while (iters--) {
+    if (pred && a_gl_rd < a_gl_end) {
+      const uint16_t* enc = reinterpret_cast<const uint16_t*>(&A[a_gl_rd]);
+      #pragma unroll
+      for (int i = 0; i < 8; i++) {
+        int4 chunk;
+        auto dec = reinterpret_cast<uint32_t*>(&chunk);
+        // We bypass the L1 cache to avoid massive amounts of memory streaming that doesn't
+        // actually help us; this brings > 2x speedup.
+        asm volatile (
+          "ld.cg.global.v4.u32 {%0, %1, %2, %3}, [%4];"
+          : "=r"(dec[0]), "=r"(dec[1]), "=r"(dec[2]), "=r"(dec[3])
+          : "l"((void*) &codebook[enc[i]])
+        );
+
+        C[a_gl_rd * 8 + i] = chunk;
+      }
+    }
+    a_gl_rd += 32;
+  }
+}
+
+
+__global__ void Code2x8Dequant(
+  const int4* __restrict__ A,
+        int4* __restrict__ C,
+  const int4* __restrict__ codebook,
+  int prob_m,
+  int prob_k,
+  const int4 codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at most 3 long, corresponds to cols.
+  const int codebook_stride // as int4
+) {
+  int a_gl_stride = prob_k / 8 / 8;
+  int a_gl_rd = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
+  bool pred = a_gl_rd < prob_m;
+
+  if (pred)
+  {
+    // advance to the correct codebook, this easy because we only multiply one column of the codebook.
+    auto codebook_size = &codebook_a_sizes.x;
+    while (a_gl_rd >= *codebook_size)
+    {
+        codebook += codebook_stride;
+        ++codebook_size;
+    }
+  }
+
+  a_gl_rd = a_gl_stride * a_gl_rd + threadIdx.x % 32;
+  int a_gl_end = a_gl_rd + a_gl_stride - threadIdx.x % 32;
+  int lane = threadIdx.x % 8;
+
+  int c_gl_stride = prob_k / 8;
+  int c_gl_wr = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
+  c_gl_wr = c_gl_stride * c_gl_wr + (threadIdx.x % 32) * 8;
+
+  extern __shared__ int4 sh[];
+  int4* sh_code = sh;
+  int4* sh_code0 = sh_code;
+  int4* sh_code1 = sh_code + 256 * 8;
+
+  for (int i = threadIdx.x; i < 2 * 256; i += blockDim.x) {
+    int4 dec = codebook[i];
+    #pragma unroll
+    for (int j = 0; j < 8; j++)
+      sh_code[8 * i + (j + lane) % 8] = dec;
+  }
+  __syncthreads();
+
+  float res = 0;
+
+  int iters = (prob_k / 8 - 1) / (8 * 32) + 1;
+  while (iters--) {
+    if (pred && a_gl_rd < a_gl_end) {
+      const uint8_t* enc = reinterpret_cast<const uint8_t*>(&A[a_gl_rd]);
+      #pragma unroll
+      for (int i = 0; i < 8; i++) {
+        int4 chunk;
+        half2* a0 = reinterpret_cast<half2*>(&sh_code0[8 * enc[2 * i + 0] + lane]);
+        half2* a1 = reinterpret_cast<half2*>(&sh_code1[8 * enc[2 * i + 1] + lane]);
+        #pragma unroll
+        for (int j = 0; j < 4; j++)
+          reinterpret_cast<half2*>(&chunk)[j] = __hadd2(a0[j], a1[j]);
+        C[a_gl_rd * 8 + i] = chunk;
+      }
+    }
+    a_gl_rd += 32;
+  }
+}
+
+inline int ceildiv(int a, int b) {
+  return (a + b - 1) / b;
+}
+
+const int THREAD_M = 16;
+
+void  code1x16_matvec_cuda(
+  const void* __restrict__ A,
+  const void* __restrict__ B,
+        void* __restrict__ C,
+  const void* __restrict__ codebook,
+  int prob_m,
+  int prob_k,
+  const int4 codebook_a_sizes,
+  const int codebook_stride
+) {
+  int sms;
+  cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, 0);
+  int waves = 0;
+  int thread_m;
+  do {
+    waves++;
+    thread_m = ceildiv(prob_m, waves * sms);
+  } while (thread_m > THREAD_M);
+
+  int blocks = ceildiv(prob_m, thread_m);
+  int threads = 32 * thread_m;
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream();
+  Code1x16MatVec<<<blocks, threads, 16*32*9, stream>>>(
+    (const int4*) A,
+    (const int4*) B,
+    (int4*) C,
+    (const int4*) codebook,
+    prob_m,
+    prob_k,
+    codebook_a_sizes,
+    codebook_stride
+  );
+}
+
+void  code2x8_matvec_cuda(
+  const void* __restrict__ A,
+  const void* __restrict__ B,
+        void* __restrict__ C,
+  const void* __restrict__ codebook,
+  int prob_m,
+  int prob_k,
+  const int4 codebook_a_sizes,
+  const int codebook_stride
+) {
+  int sms;
+  cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, 0);
+  int waves = 0;
+  int thread_m;
+  do {
+    waves++;
+    thread_m = ceildiv(prob_m, waves * sms);
+  } while (thread_m > THREAD_M);
+
+  int blocks = ceildiv(prob_m, thread_m);
+  int threads = 32 * thread_m;
+  int shared = 16 * (2 * 256 * 8 + 32 * 9);
+  cudaFuncSetAttribute(
+    Code2x8MatVec, cudaFuncAttributeMaxDynamicSharedMemorySize, shared
+  );
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream();
+  Code2x8MatVec<<<blocks, threads, shared, stream>>>(
+    (const int4*) A,
+    (const int4*) B,
+    (int4*) C,
+    (const int4*) codebook,
+    prob_m,
+    prob_k,
+    codebook_a_sizes,
+    codebook_stride
+  );
+}
+
+void code1x16_dequant_cuda(
+  const void* __restrict__ A,
+        void* __restrict__ C,
+  const void* __restrict__ codebook,
+  int prob_m,
+  int prob_k,
+  const int4 codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at most 3 long.
+  const int codebook_stride // as int4.
+) {
+  int sms;
+  cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, 0);
+  int waves = 0;
+  int thread_m;
+  do {
+    waves++;
+    thread_m = ceildiv(prob_m, waves * sms);
+  } while (thread_m > THREAD_M);
+
+  int blocks = ceildiv(prob_m, thread_m);
+  int threads = 32 * thread_m;
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream();
+  Code1x16Dequant<<<blocks, threads, 0, stream>>>(
+    (const int4*) A,
+    (int4*) C,
+    (const int4*) codebook,
+    prob_m,
+    prob_k,
+    codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at most 3 long.
+    codebook_stride // as int4.
+  );
+}
+
+// Dequantizes the code and codebook into weights.
+void  code2x8_dequant_cuda(
+  const void* __restrict__ A,
+        void* __restrict__ C,
+  const void* __restrict__ codebook,
+  int prob_m,
+  int prob_k,
+  const int4 codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at most 3 long, corresponds to cols.
+  const int codebook_stride // as int4
+) {
+  int sms;
+  cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, 0);
+  int waves = 0;
+  int thread_m;
+  do {
+    waves++;
+    thread_m = ceildiv(prob_m, waves * sms);
+  } while (thread_m > THREAD_M);
+
+  int blocks = ceildiv(prob_m, thread_m);
+  int threads = 32 * thread_m;
+  int shared = 16 * (2 * 256 * 8 + 32 * 9);
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream();
+
+  cudaFuncSetAttribute(
+    Code2x8Dequant, cudaFuncAttributeMaxDynamicSharedMemorySize, shared
+  );
+  Code2x8Dequant<<<blocks, threads, shared, stream>>>(
+    (const int4*) A,
+    (int4*) C,
+    (const int4*) codebook,
+    prob_m,
+    prob_k,
+    codebook_a_sizes,
+    codebook_stride
+  );
+}
+
+int codebook_stride(const torch::Tensor& codebooks)
+{
+  return codebooks.stride(0) * codebooks.element_size() / sizeof(int4);
+}
+
+void code1x16_matvec(
+  const torch::Tensor& A,
+  const torch::Tensor& B,
+        torch::Tensor& C,
+  const torch::Tensor& codebook,
+  const int4 codebook_a_sizes  // cumulative sizes of A spanning each codebook, at most 3 long.
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(A));
+  int prob_m = C.size(0);
+  int prob_k = B.size(0);
+
+  code1x16_matvec_cuda(
+    A.data_ptr(),
+    B.data_ptr(),
+    C.data_ptr(),
+    codebook.data_ptr(),
+    prob_m,
+    prob_k,
+    codebook_a_sizes,
+    codebook_stride(codebook)
+  );
+}
+
+torch::Tensor code1x16_matmat(
+  const torch::Tensor& input,
+  const torch::Tensor& codes,
+  const torch::Tensor& codebooks,
+  const torch::Tensor& scales,
+  const int4 codebook_a_sizes,
+  const std::optional<torch::Tensor>& bias) {
+  auto input_sizes = input.sizes();
+  auto out_features = codes.size(0) * codebooks.size(2);
+  auto flat_input = input.reshape({-1, input.size(-1)});
+  auto flat_output = torch::empty({flat_input.size(0), out_features},
+    torch::TensorOptions()
+      .dtype(input.dtype())
+      .device(input.device())
+  );
+
+  for (int i = 0; i < flat_input.size(0); ++i) {
+    auto input_vec = flat_input.index({i});
+    auto output_vec = flat_output.index({i});
+    code1x16_matvec(
+      codes.squeeze(2),
+      input_vec,
+      output_vec,
+      codebooks,
+      codebook_a_sizes
+    );
+  }
+  flat_output *= scales.flatten().unsqueeze(0);
+
+  if (bias.has_value()) {
+    flat_output += bias->unsqueeze(0);
+  }
+
+  auto output_sizes = input_sizes.vec();
+  output_sizes.pop_back();
+  output_sizes.push_back(-1);
+  auto output = flat_output.reshape(output_sizes);
+  return output;
+}
+
+void code2x8_matvec(
+  const torch::Tensor& A,
+  const torch::Tensor& B,
+        torch::Tensor& C,
+  const torch::Tensor& codebook,
+  const int4 codebook_a_sizes
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(A));
+  int prob_m = C.size(0);
+  int prob_k = B.size(0);
+  code2x8_matvec_cuda(
+    A.data_ptr(),
+    B.data_ptr(),
+    C.data_ptr(),
+    codebook.data_ptr(),
+    prob_m,
+    prob_k,
+    codebook_a_sizes,
+    2 * codebook_stride(codebook)
+  );
+}
+
+torch::Tensor code2x8_matmat(
+  const torch::Tensor& input,
+  const torch::Tensor& codes,
+  const torch::Tensor& codebooks,
+  const torch::Tensor& scales,
+  const int4 codebook_a_sizes,
+  const std::optional<torch::Tensor>& bias
+) {
+  auto input_sizes = input.sizes();
+  auto out_features = codes.size(0) * codebooks.size(2);
+  auto flat_input = input.reshape({-1, input.size(-1)});
+  auto flat_output = torch::empty({flat_input.size(0), out_features},
+    torch::TensorOptions()
+      .dtype(input.dtype())
+      .device(input.device())
+  );
+
+  for (int i = 0; i < flat_input.size(0); ++i) {
+    auto input_vec = flat_input.index({i});
+    auto output_vec = flat_output.index({i});
+    code2x8_matvec(
+      codes.squeeze(2),
+      input_vec,
+      output_vec,
+      codebooks,
+      codebook_a_sizes
+    );
+  }
+  flat_output *= scales.flatten().unsqueeze(0);
+  if (bias.has_value()) {
+    flat_output += bias->unsqueeze(0);
+  }
+
+  auto output_sizes = input_sizes.vec();
+  output_sizes.pop_back();
+  output_sizes.push_back(-1);
+  auto output = flat_output.reshape(output_sizes);
+  return output;
+}
+
+// Accumulate the partition sizes.
+int4 accumulate_sizes(const torch::Tensor& codebook_partition_sizes)
+{
+  int4 cumulative_sizes;
+  auto cumulative_size = &cumulative_sizes.x;
+  int i = 0;
+  int last = 0;
+  assert(codebook_partition_sizes.size(0) <= 4);
+  for (; i <  codebook_partition_sizes.size(0); ++i, ++cumulative_size)
+  {
+    *cumulative_size = codebook_partition_sizes[i].item<int>() + last;
+    last = *cumulative_size;
+  }
+  // fill in the rest with unreachable.
+  for (; i < 4; ++i, ++cumulative_size)
+  {
+    *cumulative_size = last*10;
+  }
+  return cumulative_sizes;
+}
+
+} // namespace aqlm
+} // namespace vllm
+
+
+torch::Tensor aqlm_gemm(
+  const torch::Tensor& input,
+  const torch::Tensor& codes,
+  const torch::Tensor& codebooks,
+  const torch::Tensor& scales,
+  const torch::Tensor& codebook_partition_sizes,
+  const std::optional<torch::Tensor>& bias
+)
+{
+  int4 cumulative_sizes = vllm::aqlm::accumulate_sizes(codebook_partition_sizes);
+
+  int const nbooks = codebooks.size(0) / codebook_partition_sizes.size(0);
+  int const entries = codebooks.size(1);
+
+  if (nbooks == 1 && entries == (1 << 16))
+  { 
+    return vllm::aqlm::code1x16_matmat(input, codes, codebooks, scales, cumulative_sizes, bias);
+  }
+  if (nbooks == 2 && entries == (1 << 8))
+  {
+    return vllm::aqlm::code2x8_matmat(input, codes, codebooks, scales, cumulative_sizes, bias);
+  }
+
+  TORCH_CHECK(false, "AQLM with ", nbooks, " codebooks and ", entries, " entries is not currently supported.")
+  return {};
+}
+
+torch::Tensor aqlm_dequant(
+  const torch::Tensor& codes,
+  const torch::Tensor& codebooks,
+  const torch::Tensor& codebook_partition_sizes
+)
+{
+  int4 cumulative_sizes = vllm::aqlm::accumulate_sizes(codebook_partition_sizes);
+
+  int const nbooks = codebooks.size(0) / codebook_partition_sizes.size(0);
+  int const entries = codebooks.size(1);
+
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(codes));
+  int rows = codes.size(1);
+  int cols = codes.size(0);
+
+  auto in_features = codes.size(1) * 8;
+  auto out_features = codes.size(0);
+
+  assert(out_features = codebook_partition_sizes.sum().item<int>());
+
+  auto weights = torch::empty({out_features, in_features},
+    torch::TensorOptions()
+      .dtype(codebooks.dtype())
+      .device(codebooks.device())
+  );
+
+  if (nbooks == 1 && entries == (1 << 16))
+  {
+    vllm::aqlm::code1x16_dequant_cuda(
+      codes.data_ptr(),
+      weights.data_ptr(),
+      codebooks.data_ptr(),
+      out_features,
+      in_features,
+      cumulative_sizes,
+      vllm::aqlm::codebook_stride(codebooks));
+
+    // if you wanted to flip to scaling the weights, (though it's 30%-ish slower and not consistent with gemv implementation.)
+    // weights *= scales.index({"...", 0, 0});
+
+     return weights;
+  }
+
+  if (nbooks == 2 && entries == (1 << 8))
+  {
+     vllm::aqlm::code2x8_dequant_cuda(
+        codes.data_ptr(), 
+        weights.data_ptr(), 
+        codebooks.data_ptr(), 
+        out_features,
+        in_features, 
+        cumulative_sizes, 
+        vllm::aqlm::codebook_stride(codebooks));
+
+    // if you wanted to flip to scaling the weights, (though it's 30%-ish slower and not consistent with gemv implementation)
+    // weights *= scales.index({"...", 0, 0});
+
+     return weights;
+  }
+
+  TORCH_CHECK(false, "AQLM with ", nbooks, " codebooks and ", entries, " entries is not currently supported.")
+  return {};
+}

examples/aqlm_example.py

+import argparse
+
+from vllm import LLM, SamplingParams
+
+
+def main():
+
+    parser = argparse.ArgumentParser(description='AQLM examples')
+
+    parser.add_argument('--model',
+                        '-m',
+                        type=str,
+                        default=None,
+                        help='model path, as for HF')
+    parser.add_argument('--choice',
+                        '-c',
+                        type=int,
+                        default=0,
+                        help='known good models by index, [0-4]')
+    parser.add_argument('--tensor_parallel_size',
+                        '-t',
+                        type=int,
+                        default=1,
+                        help='tensor parallel size')
+
+    args = parser.parse_args()
+
+    models = [
+        "ISTA-DASLab/Llama-2-7b-AQLM-2Bit-1x16-hf",
+        "ISTA-DASLab/Llama-2-7b-AQLM-2Bit-2x8-hf",
+        "ISTA-DASLab/Llama-2-13b-AQLM-2Bit-1x16-hf",
+        "ISTA-DASLab/Mixtral-8x7b-AQLM-2Bit-1x16-hf",
+        "BlackSamorez/TinyLlama-1_1B-Chat-v1_0-AQLM-2Bit-1x16-hf",
+    ]
+
+    model = LLM(args.model if args.model is not None else models[args.choice],
+                tensor_parallel_size=args.tensor_parallel_size)
+
+    sampling_params = SamplingParams(max_tokens=100, temperature=0)
+    outputs = model.generate("Hello my name is",
+                             sampling_params=sampling_params)
+    print(outputs[0].outputs[0].text)
+
+
+if __name__ == '__main__':
+    main()

tests/models/test_aqlm.py

+"""Compare the outputs of a AQLM model between vLLM and HF Transformers
+
+Run `pytest tests/models/test_aqlm.py`.
+"""
+
+import pytest
+import torch
+
+from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
+
+capability = torch.cuda.get_device_capability()
+capability = capability[0] * 10 + capability[1]
+aqlm_not_supported = (capability <
+                      QUANTIZATION_METHODS["aqlm"].get_min_capability())
+
+# In this test we hardcode prompts and generations for the model so we don't
+# need to require the AQLM package as a dependency
+example_prompts = [
+    'vLLM is a high-throughput and memory-efficient inference and serving '
+    'engine for LLMs.\n',
+    'Briefly describe the major milestones in the development of artificial '
+    'intelligence from 1950 to 2020.\n',
+    'Compare and contrast artificial intelligence with human intelligence in '
+    'terms of processing information.\n',
+    'Describe the basic components of a neural network and how it can be '
+    'trained.\n',
+    'Write a short story about a robot that dreams for the first time.\n',
+    'Analyze the impact of the COVID-19 pandemic on global economic structures '
+    'and future business models.\n',
+    'Explain the cultural significance of the Mona Lisa painting, and how its '
+    'perception might vary in Western versus Eastern societies.\n',
+    "Translate the following English sentence into Japanese, French, and "
+    "Swahili: 'The early bird catches the worm.'\n"
+]
+
+# These ground truth generations were generated using `transformers==4.38.1
+# aqlm==1.1.0 torch==2.2.0`
+# and the below code:
+# ```python
+# from transformers import AutoTokenizer, AutoModelForCausalLM
+# model_id = "ISTA-DASLab/Llama-2-7b-AQLM-2Bit-1x16-hf"
+# quantized_model = AutoModelForCausalLM.from_pretrained(model_id,
+# torch_dtype="auto", device_map="cuda").cuda()
+# tokenizer = AutoTokenizer.from_pretrained(model_id)
+# outputs = []
+# for prompt in example_prompts:
+#     input_ids = tokenizer(prompt, return_tensors="pt")["input_ids"].to("cuda")
+#     hf_outputs = quantized_model.generate(input_ids, max_new_tokens=32)
+#     outputs.append(tokenizer.decode(hf_outputs[0][input_ids.shape[1]:]))
+# print(outputs)
+# ```
+ground_truth_generations = [
+    '\n### Features\n\n- **High-throughput**: v',
+    'The major milestones in the development of artificial intelligence from '
+    '195',
+    'Compare and contrast artificial intelligence with human intelligence in '
+    'terms of processing information. The',
+    'Explain the difference between supervised and unsupervised learning.'
+    '\nExplain',
+    'Write a short story about a robot that dreams for the first time. The',
+    'Analyze the impact of the COVID-19 pandemic on global economic',
+    'The Mona Lisa is a painting by Leonardo da Vinci, and it',
+    'The early bird catches the worm.\nThe early bird catches the'
+]
+
+
+@pytest.mark.skipif(aqlm_not_supported,
+                    reason="AQLM is not supported on this GPU type.")
+@pytest.mark.parametrize("model", ["ISTA-DASLab/Llama-2-7b-AQLM-2Bit-1x16-hf"])
+@pytest.mark.parametrize("dtype", ["half"])
+@pytest.mark.parametrize("max_tokens", [16])
+@pytest.mark.parametrize("num_logprobs", [1])
+def test_models(
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+    num_logprobs: int,
+) -> None:
+
+    vllm_model = vllm_runner(model, dtype=dtype)
+    vllm_outputs = vllm_model.generate_greedy_logprobs(example_prompts,
+                                                       max_tokens,
+                                                       num_logprobs)
+
+    # loop through the prompts to compare against the ground truth generations
+    for prompt_idx in range(len(example_prompts)):
+        vllm_output_ids, vllm_output_str, vllm_logprobs = vllm_outputs[
+            prompt_idx]
+
+        print("Prompt:          ", repr(example_prompts[prompt_idx]))
+        print("Reference output:", repr(ground_truth_generations[prompt_idx]))
+        print("Output output:   ", repr(vllm_output_str))
+        assert vllm_output_str == ground_truth_generations[prompt_idx]

vllm/model_executor/layers/linear.py

@@ -31,7 +31,7 @@ class LinearMethodBase(ABC):
     @abstractmethod
     def create_weights(self, layer: torch.nn.Module,
                        input_size_per_partition: int,
-                       output_size_per_partition: int, input_size: int,
+                       output_partition_sizes: List[int], input_size: int,
                        output_size: int, params_dtype: torch.dtype,
                        **extra_weight_attrs):
         """Create weights for a linear layer.
@@ -70,9 +70,10 @@ class UnquantizedLinearMethod(LinearMethodBase):
 
     def create_weights(self, layer: torch.nn.Module,
                        input_size_per_partition: int,
-                       output_size_per_partition: int, input_size: int,
+                       output_partition_sizes: List[int], input_size: int,
                        output_size: int, params_dtype: torch.dtype,
                        **extra_weight_attrs):
+        output_size_per_partition = sum(output_partition_sizes)
         weight = Parameter(torch.empty(output_size_per_partition,
                                        input_size_per_partition,
                                        dtype=params_dtype),
@@ -127,7 +128,7 @@ class ReplicatedLinear(torch.nn.Module):
             linear_method = UnquantizedLinearMethod()
         self.linear_method = linear_method
         self.linear_method.create_weights(self, self.input_size,
-                                          self.output_size, self.input_size,
+                                          [self.output_size], self.input_size,
                                           self.output_size, self.params_dtype)
         if bias:
             self.bias = Parameter(
@@ -161,6 +162,8 @@ class ColumnParallelLinear(torch.nn.Module):
                        skip adding bias but instead return it.
         params_dtype: Data type for the parameters.
         linear_method: (Maybe quantized) linear method.
+        output_sizes: list of output sizes packed into one output, like for QKV
+                       the list would be size 3.
     """
 
     def __init__(
@@ -172,6 +175,7 @@ class ColumnParallelLinear(torch.nn.Module):
         skip_bias_add: bool = False,
         params_dtype: Optional[torch.dtype] = None,
         linear_method: Optional[LinearMethodBase] = None,
+        output_sizes: Optional[List[int]] = None,
     ):
         super().__init__()
 
@@ -188,10 +192,12 @@ class ColumnParallelLinear(torch.nn.Module):
         self.params_dtype = params_dtype
         if linear_method is None:
             linear_method = UnquantizedLinearMethod()
+        if output_sizes is None:
+            output_sizes = [output_size]
         self.linear_method = linear_method
         self.linear_method.create_weights(self,
                                           self.input_size,
-                                          self.output_size_per_partition,
+                                          [x // tp_size for x in output_sizes],
                                           self.input_size,
                                           self.output_size,
                                           self.params_dtype,
@@ -268,14 +274,17 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
         tp_size = get_tensor_model_parallel_world_size()
         assert all(output_size % tp_size == 0 for output_size in output_sizes)
         super().__init__(input_size, sum(output_sizes), bias, gather_output,
-                         skip_bias_add, params_dtype, linear_method)
+                         skip_bias_add, params_dtype, linear_method,
+                         self.output_sizes)
 
     def weight_loader(self,
                       param: Parameter,
                       loaded_weight: torch.Tensor,
                       loaded_shard_id: Optional[int] = None):
+
         param_data = param.data
         output_dim = getattr(param, "output_dim", None)
+        is_metadata = getattr(param, "is_metadata", False)
         if loaded_shard_id is None:
             # Loaded weight is already packed.
             if output_dim is None:
@@ -328,6 +337,11 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
             start_idx = tp_rank * shard_size
             loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                  shard_size)
+        elif is_metadata:
+            # metadata indicates fixed size concatenated along dim 0
+            shard_size = loaded_weight.shape[0]
+            shard_offset = loaded_shard_id * shard_size
+            param_data = param_data.narrow(0, shard_offset, shard_size)
         else:
             ignore_warning = getattr(param, "ignore_warning", False)
             if not ignore_warning:
@@ -393,8 +407,14 @@ class QKVParallelLinear(ColumnParallelLinear):
         input_size = self.hidden_size
         output_size = (self.num_heads +
                        2 * self.num_kv_heads) * tp_size * self.head_size
+        output_sizes = [
+            self.num_heads * tp_size * self.head_size,
+            self.num_kv_heads * tp_size * self.head_size,
+            self.num_kv_heads * tp_size * self.head_size
+        ]
+
         super().__init__(input_size, output_size, bias, False, skip_bias_add,
-                         params_dtype, linear_method)
+                         params_dtype, linear_method, output_sizes)
 
     def weight_loader(self,
                       param: Parameter,
@@ -402,6 +422,7 @@ class QKVParallelLinear(ColumnParallelLinear):
                       loaded_shard_id: Optional[str] = None):
         param_data = param.data
         output_dim = getattr(param, "output_dim", None)
+        is_metadata = getattr(param, "is_metadata", False)
 
         if loaded_shard_id is None:
             # Loaded weight is already packed.
@@ -469,6 +490,12 @@ class QKVParallelLinear(ColumnParallelLinear):
             start_idx = shard_id * shard_size
             loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                  shard_size)
+        elif is_metadata:
+            # metadata indicates fixed size concatenated along dim 0
+            shard_size = loaded_weight.shape[0]
+            shard_index = ["q", "k", "v"].index(loaded_shard_id)
+            param_data = param_data.narrow(0, shard_index * shard_size,
+                                           shard_size)
         else:
             ignore_warning = getattr(param, "ignore_warning", False)
             if not ignore_warning:
@@ -536,7 +563,7 @@ class RowParallelLinear(torch.nn.Module):
         self.linear_method = linear_method
         self.linear_method.create_weights(self,
                                           self.input_size_per_partition,
-                                          self.output_size,
+                                          [self.output_size],
                                           self.input_size,
                                           self.output_size,
                                           self.params_dtype,

vllm/model_executor/layers/quantization/__init__.py

 from typing import Type
 
+from vllm.model_executor.layers.quantization.aqlm import AQLMConfig
 from vllm.model_executor.layers.quantization.awq import AWQConfig
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
@@ -9,6 +10,7 @@ from vllm.model_executor.layers.quantization.marlin import MarlinConfig
 from vllm.model_executor.layers.quantization.squeezellm import SqueezeLLMConfig
 
 QUANTIZATION_METHODS = {
+    "aqlm": AQLMConfig,
     "awq": AWQConfig,
     "fp8": FP8Config,
     "gptq": GPTQConfig,

vllm/model_executor/layers/quantization/aqlm.py

+# Supports AQLM compression, see https://github.com/Vahe1994/AQLM
+# and https://arxiv.org/pdf/2401.06118.pdf
+
+import math
+from typing import Any, Dict, List, Optional
+
+import torch
+import torch.nn.functional as F
+from torch.nn.parameter import Parameter
+
+from vllm._C import ops
+from vllm.model_executor.layers.linear import (LinearMethodBase,
+                                               set_weight_attrs)
+from vllm.model_executor.layers.quantization.base_config import (
+    QuantizationConfig)
+
+
+def get_int_dtype(nbits: int) -> torch.dtype:
+    if nbits <= 8:
+        return torch.int8
+    if nbits <= 16:
+        return torch.int16
+    if nbits <= 32:
+        return torch.int32
+    if nbits <= 64:
+        return torch.int64
+    raise ValueError(f"No dtype available for {nbits}-bit codebooks")
+
+
+@torch.inference_mode()
+def unpack_int_data(data: torch.IntTensor, nbits: int) -> torch.IntTensor:
+    return data.to(torch.int64) % (2**nbits)
+
+
+def dequantize_weight(codes: torch.Tensor,
+                      codebooks: torch.Tensor,
+                      scales: Optional[torch.Tensor] = None) -> torch.Tensor:
+    """
+    Decode float weights from quantization codes. Differentiable.
+    :param codes: tensor of integer quantization codes, shape 
+        [*dims, num_out_groups, num_in_groups, num_codebooks]
+    :param codebooks: tensor of vectors for each quantization code, 
+        [num_codebooks, codebook_size, out_group_size, in_group_size]
+    :param scales: weight will be multiplied by this factor, must be 
+        broadcastble with 
+        [*dims, out_groups, num_in_groups, out_group_size, in_group_size]
+    :return: reconstructed weight tensor of shape 
+        [*dims, num_in_groups*group_size]
+    """
+    num_out_groups, num_in_groups, num_codebooks = codes.shape[-3:]
+    num_codebooks, codebook_size, out_group_size, in_group_size = \
+        codebooks.shape
+    out_features = num_out_groups * out_group_size
+    in_features = num_in_groups * in_group_size
+    codebook_offsets = torch.arange(
+        0, num_codebooks * codebook_size, codebook_size,
+        device=codes.device)  # shape: [num_codebooks]
+    reconstructed_weight_flat = F.embedding_bag(
+        codes.flatten(0, -2) + codebook_offsets,
+        codebooks.flatten(0, 1).flatten(-2, -1),
+        mode="sum"
+    )  # [prod(dims) * num_out_groups * num_in_groups, out_group_size
+    # * in_group_size]
+
+    reconstructed_weight_groupwise = reconstructed_weight_flat.view(
+        list(codes.shape[:-3]) +
+        [num_out_groups, num_in_groups, out_group_size, in_group_size])
+    if scales is not None:
+        reconstructed_weight_groupwise = reconstructed_weight_groupwise.mul(
+            scales)
+    return reconstructed_weight_groupwise.swapaxes(
+        -3, -2).reshape(list(codes.shape[:-3]) + [out_features, in_features])
+
+
+def dequantize_gemm(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+    dequantized_weight = dequantize_weight(
+        unpack_int_data(codes, codebooks.shape[1].bit_length() - 1),
+        codebooks,
+        scales,
+    )
+    return F.linear(input, dequantized_weight, bias)
+
+
+# Generic dequantization, slow but flexible.
+def generic_dequantize_gemm(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    output_partition_sizes: torch.IntTensor,
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+    output_shape = input.shape[:-1] + (scales.shape[0], )
+    output = torch.empty(output_shape, dtype=input.dtype, device=input.device)
+    num_outputs = len(output_partition_sizes)
+
+    # break the inputs and codebooks apart then combine the outputs.
+    # Surprisingly (to me) this is faster than doing 3 de-quants and 1 big
+    # multiply at the end.
+    num_codebooks = codebooks.shape[0] // num_outputs
+    assert (scales.shape[0] == codes.shape[0])
+    assert (sum(output_partition_sizes) == scales.shape[0])
+    output_offset = 0
+    codebooks_offset = 0
+    for output_size in output_partition_sizes:
+        shard_output = dequantize_gemm(
+            input, codes.narrow(0, output_offset, output_size),
+            codebooks.narrow(0, codebooks_offset, num_codebooks),
+            scales.narrow(0, output_offset, output_size), None
+            if bias is None else bias.narrow(0, output_offset, output_size))
+
+        output_slice = output.narrow(-1, output_offset, output_size)
+        assert (output_slice.shape == shard_output.shape)
+        output_slice.copy_(shard_output)
+        output_offset += output_size
+        codebooks_offset += num_codebooks
+    return output
+
+
+# Optimized dequnantize/decompression kernels, supports 1x16 and 2x8
+# at 6 and 9 times faster than the generic version above, respectively.
+def optimized_dequantize_gemm(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    output_partition_sizes: torch.IntTensor,
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+    weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
+
+    if bias is None:
+        # scaling the output is fastest, so we do that when possible.
+        output = F.linear(input, weights, bias)
+        orig_shape = output.shape
+        flattened_output = output.view(-1, output.size(-1))
+        f_scales = scales.view(-1, scales.shape[0])
+        b_scales = f_scales.expand(flattened_output.shape[0], -1)
+        flattened_output *= b_scales
+        return output.view(orig_shape)
+    else:
+        b_scales = scales.view(scales.shape[:-3] + (-1, )).expand(
+            -1, weights.shape[1])
+        weights *= b_scales
+        return F.linear(input, weights, bias)
+
+
+class AQLMConfig(QuantizationConfig):
+    """Config class for AQLM.
+
+    Reference: https://github.com/Vahe1994/AQLM
+    """
+
+    def __init__(
+        self,
+        in_group_size: int,
+        nbits_per_codebook: int,
+        num_codebooks: int,
+        out_group_size: int,
+    ) -> None:
+        self.in_group_size = in_group_size
+        self.nbits_per_codebook = nbits_per_codebook
+        self.num_codebooks = num_codebooks
+        self.out_group_size = out_group_size
+
+        # out_group_size > 1 is untested, and probably won't work as-is.
+        assert (self.out_group_size == 1)
+        self.pack_factor = (self.in_group_size * self.out_group_size)
+
+    def __repr__(self) -> str:
+        return (f"AQLMConfig(in_group_size={self.in_group_size}, "
+                f"nbits_per_codebook={self.nbits_per_codebook}, "
+                f"num_codebooks={self.num_codebooks}, "
+                f"out_group_size={self.out_group_size})")
+
+    @classmethod
+    def get_name(cls) -> str:
+        return "aqlm"
+
+    @classmethod
+    def get_supported_act_dtypes(cls) -> List[torch.dtype]:
+        return [torch.half]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 70
+
+    @classmethod
+    def get_config_filenames(cls) -> List[str]:
+        return []  # no extra configs.
+
+    @classmethod
+    def from_config(cls, config: Dict[str, Any]) -> "AQLMConfig":
+        in_group_size = cls.get_from_keys(config, ["in_group_size"])
+        nbits_per_codebook = cls.get_from_keys(config, ["nbits_per_codebook"])
+        num_code_books = cls.get_from_keys(config, ["num_codebooks"])
+        out_group_size = cls.get_from_keys(config, ["out_group_size"])
+        return cls(in_group_size, nbits_per_codebook, num_code_books,
+                   out_group_size)
+
+    def get_linear_method(self) -> "AQLMLinearMethod":
+        return AQLMLinearMethod(self)
+
+    def get_scaled_act_names(self) -> List[str]:
+        return []
+
+
+class AQLMLinearMethod(LinearMethodBase):
+    """Linear method for AQLM.
+
+    Args:
+        quant_config: The AQLM quantization config.
+    """
+
+    def __init__(self, quant_config: AQLMConfig):
+        self.quant_config = quant_config
+
+    def create_weights(self, layer: torch.nn.Module,
+                       input_size_per_partition: int,
+                       output_partition_sizes: List[int], input_size: int,
+                       output_size: int, params_dtype: torch.dtype,
+                       **extra_weight_attrs):
+        del output_size  # Unused.
+        del input_size  # Unused.
+
+        if params_dtype != torch.half:
+            raise ValueError("Only half is currently supported by aqlm")
+        if input_size_per_partition % self.quant_config.in_group_size != 0:
+            raise ValueError(
+                "The input size is not aligned with the quantized "
+                "weight shape. This can be caused by too large "
+                "tensor parallel size.")
+
+        output_size_per_partition = sum(output_partition_sizes)
+        if output_size_per_partition % self.quant_config.out_group_size != 0:
+            raise ValueError(
+                "The output size is not aligned with the quantized "
+                "weight shape. This can be caused by too large "
+                "tensor parallel size.")
+
+        codes = Parameter(
+            torch.empty(
+                # There could actually be two pack factors, one along input and
+                # one along output, but we don't currently support
+                # out_group_size, and only the one along output needs to be
+                # marked with "packed_dim" in order for QKVLinear to work.
+                output_size_per_partition,
+                input_size_per_partition // self.quant_config.pack_factor,
+                self.quant_config.num_codebooks,
+                dtype=get_int_dtype(self.quant_config.nbits_per_codebook),
+            ),
+            requires_grad=False,
+        )
+
+        set_weight_attrs(
+            codes,
+            {
+                "input_dim": 1,
+                "output_dim": 0,
+                "packed_dim": 1,
+                "pack_factor": self.quant_config.pack_factor,
+            },
+        )
+
+        codebooks = Parameter(
+            torch.empty(
+                self.quant_config.num_codebooks * len(output_partition_sizes),
+                2**self.quant_config.nbits_per_codebook,
+                self.quant_config.out_group_size,
+                self.quant_config.in_group_size,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        set_weight_attrs(
+            codebooks,
+            {
+                # metadata indicates fixed size concatenated along dim 0
+                "is_metadata":
+                True,
+                "output_partition_sizes":
+                torch.tensor(output_partition_sizes, device='cpu'),
+            },
+        )
+
+        scales = Parameter(
+            torch.empty(
+                (
+                    output_size_per_partition //
+                    self.quant_config.out_group_size,
+                    1,
+                    1,
+                    1,
+                ),
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        set_weight_attrs(
+            scales,
+            {
+                "output_dim": 0,
+                "packed_dim": 0,
+                "pack_factor": self.quant_config.out_group_size
+            },
+        )
+
+        layer.register_parameter("codes", codes)
+        set_weight_attrs(codes, extra_weight_attrs)
+        layer.register_parameter("codebooks", codebooks)
+        set_weight_attrs(codebooks, extra_weight_attrs)
+        layer.register_parameter("scales", scales)
+        set_weight_attrs(scales, extra_weight_attrs)
+
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        codebooks = layer.codebooks
+        codes = layer.codes
+        scales = layer.scales
+        output_partition_sizes = getattr(codebooks, "output_partition_sizes",
+                                         None)
+
+        nbooks = codes.shape[2]
+        ingroups = codebooks.shape[3]
+        outgroups = codebooks.shape[2]
+        bits = codebooks.shape[1]
+
+        # We support these formats with dedicated gemm and decompression
+        # kernels.
+        if ingroups == 8 and outgroups == 1 and (
+            (bits == 256 and nbooks == 2) or (bits == 65536 and nbooks == 1)):
+
+            # thresholds determined by timings on an A6000, one GPU
+            use_gemv = math.prod(x.shape[:-1]) <= 6
+
+            return ops.aqlm_gemm(
+                x,
+                codes,
+                codebooks,
+                scales,
+                output_partition_sizes,
+                bias,
+            ) if use_gemv else optimized_dequantize_gemm(
+                x,
+                codes,
+                codebooks,
+                scales,
+                output_partition_sizes,
+                bias,
+            )
+
+        # fall back all unoptimized formats
+        return generic_dequantize_gemm(
+            x,
+            codes,
+            codebooks,
+            scales,
+            output_partition_sizes,
+            bias,
+        )

vllm/model_executor/layers/quantization/awq.py

@@ -81,7 +81,7 @@ class AWQLinearMethod(LinearMethodBase):
 
     def create_weights(self, layer: torch.nn.Module,
                        input_size_per_partition: int,
-                       output_size_per_partition: int, input_size: int,
+                       output_partition_sizes: List[int], input_size: int,
                        output_size: int, params_dtype: torch.dtype,
                        **extra_weight_attrs):
         if input_size_per_partition % self.quant_config.group_size != 0:
@@ -89,6 +89,8 @@ class AWQLinearMethod(LinearMethodBase):
                 "The input size is not aligned with the quantized "
                 "weight shape. This can be caused by too large "
                 "tensor parallel size.")
+
+        output_size_per_partition = sum(output_partition_sizes)
         if output_size_per_partition % self.quant_config.pack_factor != 0:
             raise ValueError(
                 "The output size is not aligned with the quantized "

vllm/model_executor/layers/quantization/gptq.py

@@ -91,7 +91,7 @@ class GPTQLinearMethod(LinearMethodBase):
         self,
         layer: torch.nn.Module,
         input_size_per_partition: int,
-        output_size_per_partition: int,
+        output_partition_sizes: List[int],
         input_size: int,
         output_size: int,
         params_dtype: torch.dtype,
@@ -103,6 +103,7 @@ class GPTQLinearMethod(LinearMethodBase):
                 "The input size is not aligned with the quantized "
                 "weight shape. This can be caused by too large "
                 "tensor parallel size.")
+        output_size_per_partition = sum(output_partition_sizes)
         if (output_size_per_partition % self.quant_config.pack_factor.numerator
                 != 0):
             raise ValueError(

vllm/model_executor/layers/quantization/marlin.py

@@ -93,7 +93,7 @@ class MarlinLinearMethod(LinearMethodBase):
         self,
         layer: torch.nn.Module,
         input_size_per_partition: int,
-        output_size_per_partition: int,
+        output_partition_sizes: List[int],
         input_size: int,
         output_size: int,
         params_dtype: torch.dtype,
@@ -106,6 +106,7 @@ class MarlinLinearMethod(LinearMethodBase):
                 f"The params dtype must be float16, but got {params_dtype}")
 
         # Validate output_size_per_partition
+        output_size_per_partition = sum(output_partition_sizes)
         if output_size_per_partition % self.quant_config.min_n_threads != 0:
             raise ValueError(
                 f"Weight output_size_per_partition = "

vllm/model_executor/layers/quantization/squeezellm.py

@@ -70,7 +70,7 @@ class SqueezeLLMLinearMethod(LinearMethodBase):
 
     def create_weights(self, layer: torch.nn.Module,
                        input_size_per_partition: int,
-                       output_size_per_partition: int, input_size: int,
+                       output_partition_sizes: List[int], input_size: int,
                        output_size: int, params_dtype: torch.dtype,
                        **extra_weight_attrs):
         if input_size_per_partition % self.quant_config.pack_factor != 0:
@@ -78,6 +78,8 @@ class SqueezeLLMLinearMethod(LinearMethodBase):
                 "The input size is not aligned with the quantized "
                 "weight shape. This can be caused by too large "
                 "tensor parallel size.")
+
+        output_size_per_partition = sum(output_partition_sizes)
         qweight = Parameter(
             torch.empty(
                 input_size_per_partition // self.quant_config.pack_factor,