Skip to content

GitLab

"vllm/vscode:/vscode.git/clone" did not exist on "c60f8e3b49eced1a17ba0e11da3f8c107b309df9"
Commit 99324e25 authored by zhuwenwen's avatar zhuwenwen
Browse files

Merge tag 'v0.9.2' into v0.9.2-ori

parents cc7f22a8 a5dd03c1
Hide whitespace changes
Inline Side-by-side
Showing with 1723 additions and 156 deletions
csrc/quickreduce/base.h 0 → 100644
View file @ 99324e25
#pragma once
#include <cstdint>
#include <hip/hip_runtime.h>
#include <hip/hip_fp16.h>
#include <hip/hip_bf16.h>
#define __quickreduce_device_inline__ __device__ __forceinline__
#define __quickreduce_launch_bounds_two_shot__ __launch_bounds__(256, 4)
#define __quickreduce_launch_bounds_one_shot__ __launch_bounds__(512, 4)
namespace quickreduce {
typedef __hip_bfloat16 nv_bfloat16;
typedef __hip_bfloat162 nv_bfloat162;
using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
// Setup acquire-release semantics for vector memory reads (mubuf instruction)
// as per architecture.
#if defined(__gfx942__)
// CDNA3: Scope bits sc0, sc1
#define MUBUF_ACQUIRE 16
#define MUBUF_RELEASE 16
#elif (defined(__gfx908__) || defined(__gfx90a__))
// CDNA1 and CDNA2 - glc bit
#define MUBUF_ACQUIRE 1
#define MUBUF_RELEASE 0
#endif
static constexpr int kNegOne = 0xBC00BC00; // {-1, -1}, fp16x2_t
// Number of atoms (4xf16x2_t) processed by a single thread
static constexpr int kAtoms = 8;
// We use a workgroup of 256 threads
static constexpr int kBlockSize = 256;
static constexpr int kAtomStride = kBlockSize;
// Size and atom stride of source/destination data that the block will
// process.
// Workgroup scope = Tile = (256 threads x 8 atoms x 16B)
static constexpr int kTileSize = kBlockSize * kAtoms * sizeof(int32x4_t);
// Max number of blocks. 304 CUs on MI300
static constexpr int kMaxNumBlocks = 304 * 4;
// Standard CDNA wavefront size.
static constexpr int kWavefront = 64;
// 256 thread, 4 wavefronts.
static dim3 constexpr kBlockTwoShot = {kWavefront, kBlockSize / kWavefront, 1};
// Number of threads in a group for quantization
// It corresponds to 32 F16 elements in quantization block
static constexpr int kThreadGroupSize = 8;
// Methods
__quickreduce_device_inline__ __host__ unsigned long divceil(unsigned long x,
unsigned long y) {
return ((x + y - 1) / y);
}
union BufferResource {
__quickreduce_device_inline__ constexpr BufferResource()
: config(0x00020000U) {}
__quickreduce_device_inline__ constexpr BufferResource(void* buffer_address,
uint32_t buffer_size)
: address(buffer_address), range(buffer_size), config(0x00020000U) {}
int32x4_t descriptor;
struct {
void* address; // 8B, out of which first 48b is address, and 16b is stride
// (unused)
uint32_t range; // Byte range for the buffer resource
uint32_t config; // Constant, DFMT=32b
};
};
__quickreduce_device_inline__ static int32x4_t buffer_load_dwordx4(
int32x4_t srsrc, int32_t voffset, int32_t soffset,
int32_t aux) __asm("llvm.amdgcn.raw.buffer.load.v4i32");
__quickreduce_device_inline__ static void buffer_store_dwordx4(
int32x4_t data, int32x4_t srsrc, int32_t voffset, int32_t soffset,
int32_t aux) __asm("llvm.amdgcn.raw.buffer.store.v4i32");
__quickreduce_device_inline__ static void set_fp16_ovfl(bool const value) {
#if defined(__gfx942__)
if (value) {
asm volatile("s_setreg_imm32_b32 0xdc1, 1;" ::);
} else {
asm volatile("s_setreg_imm32_b32 0xdc1, 0;" ::);
}
#endif
}
union bf162_int_union {
int i;
nv_bfloat162 bf2;
};
template <typename T>
__quickreduce_device_inline__ void packed_assign_add(int32x4_t* A,
int32x4_t* B);
template <>
__quickreduce_device_inline__ void packed_assign_add<half>(int32x4_t* A,
int32x4_t* B) {
int32x4_t& tR_fragment = A[0];
int32x4_t& tA_fragment = B[0];
asm volatile("v_pk_add_f16 %0, %1, %2"
: "=v"(tR_fragment[0])
: "v"(tR_fragment[0]), "v"(tA_fragment[0]));
asm volatile("v_pk_add_f16 %0, %1, %2"
: "=v"(tR_fragment[1])
: "v"(tR_fragment[1]), "v"(tA_fragment[1]));
asm volatile("v_pk_add_f16 %0, %1, %2"
: "=v"(tR_fragment[2])
: "v"(tR_fragment[2]), "v"(tA_fragment[2]));
asm volatile("v_pk_add_f16 %0, %1, %2"
: "=v"(tR_fragment[3])
: "v"(tR_fragment[3]), "v"(tA_fragment[3]));
}
template <>
__quickreduce_device_inline__ void packed_assign_add<nv_bfloat16>(
int32x4_t* A, int32x4_t* B) {
nv_bfloat162* tA = reinterpret_cast<nv_bfloat162*>(A);
nv_bfloat162* tB = reinterpret_cast<nv_bfloat162*>(B);
#pragma unroll
for (int i = 0; i < 4; i++) {
tA[i] = __hadd2(tA[i], tB[i]);
}
}
template <typename T>
__quickreduce_device_inline__ int packed_max(int a, int b);
template <>
__quickreduce_device_inline__ int packed_max<half>(int a, int b) {
int result;
asm volatile("v_pk_max_f16 %0, %1, %2" : "=v"(result) : "v"(a), "v"(b));
return result;
}
template <>
__quickreduce_device_inline__ int packed_max<nv_bfloat16>(int a, int b) {
bf162_int_union A, B, R;
A.i = a;
B.i = b;
R.bf2 = __hmax2(A.bf2, B.bf2);
return R.i;
}
template <typename T>
__quickreduce_device_inline__ int packed_min(int a, int b);
template <>
__quickreduce_device_inline__ int packed_min<half>(int a, int b) {
int result;
asm volatile("v_pk_min_f16 %0, %1, %2" : "=v"(result) : "v"(a), "v"(b));
return result;
}
template <>
__quickreduce_device_inline__ int packed_min<nv_bfloat16>(int a, int b) {
bf162_int_union A, B, R;
A.i = a;
B.i = b;
R.bf2 = __hmin2(A.bf2, B.bf2);
return R.i;
}
template <typename T>
__quickreduce_device_inline__ int packed_abs_max(int a, int b);
template <>
__quickreduce_device_inline__ int packed_abs_max<half>(int a, int b) {
half2 wmaxh2 = __builtin_bit_cast(half2, a);
half2 wminh2 = __builtin_bit_cast(half2, b);
half2 wblockmaxh2;
wblockmaxh2.x =
__hgt(__habs(wmaxh2.x), __habs(wminh2.x)) ? wmaxh2.x : wminh2.x;
wblockmaxh2.y =
__hgt(__habs(wmaxh2.y), __habs(wminh2.y)) ? wmaxh2.y : wminh2.y;
return __builtin_bit_cast(int, wblockmaxh2);
}
template <>
__quickreduce_device_inline__ int packed_abs_max<nv_bfloat16>(int a, int b) {
bf162_int_union A, B, R;
A.i = a;
B.i = b;
R.bf2.x = __hgt(__habs(A.bf2.x), __habs(B.bf2.x)) ? A.bf2.x : B.bf2.x;
R.bf2.y = __hgt(__habs(A.bf2.y), __habs(B.bf2.y)) ? A.bf2.y : B.bf2.y;
return R.i;
}
template <typename T>
__quickreduce_device_inline__ int packed_add(int a, int b);
template <>
__quickreduce_device_inline__ int packed_add<half>(int a, int b) {
int result;
asm volatile("v_pk_add_f16 %0, %1, %2" : "=v"(result) : "v"(a), "v"(b));
return result;
}
template <>
__quickreduce_device_inline__ int packed_add<nv_bfloat16>(int a, int b) {
bf162_int_union A, B, R;
A.i = a;
B.i = b;
R.bf2 = __hadd2(A.bf2, B.bf2);
return R.i;
}
template <>
__quickreduce_device_inline__ int packed_add<int16_t>(int a, int b) {
int result;
asm volatile("v_pk_add_i16 %0, %1, %2" : "=v"(result) : "v"(a), "v"(b));
return result;
}
template <typename T>
__quickreduce_device_inline__ int packed_sub(int a, int b);
template <>
__quickreduce_device_inline__ int packed_sub<half>(int a, int b) {
int result;
// MI300 lacks packed fp16 sub instruction. So we do -1 * min + max
asm volatile("v_pk_fma_f16 %0, %1, %2 %3"
: "=v"(result)
: "v"(kNegOne), "v"(b), "v"(a));
return result;
}
template <>
__quickreduce_device_inline__ int packed_sub<nv_bfloat16>(int a, int b) {
bf162_int_union A, B, R;
A.i = a;
B.i = b;
R.bf2 = __hsub2(A.bf2, B.bf2);
return R.i;
}
template <typename T>
__quickreduce_device_inline__ int packed_mul(int a, int b);
template <>
__quickreduce_device_inline__ int packed_mul<half>(int a, int b) {
int result;
asm volatile("v_pk_mul_f16 %0, %1, %2" : "=v"(result) : "v"(a), "v"(b));
return result;
}
template <>
__quickreduce_device_inline__ int packed_mul<nv_bfloat16>(int a, int b) {
nv_bfloat162* tA = reinterpret_cast<nv_bfloat162*>(&a);
nv_bfloat162* tB = reinterpret_cast<nv_bfloat162*>(&b);
nv_bfloat162 tR = __hmul2(*tA, *tB);
return *(reinterpret_cast<int*>(&tR));
}
template <typename T>
__quickreduce_device_inline__ int packed_rcp(int a);
template <>
__quickreduce_device_inline__ int packed_rcp<half>(int a) {
return __builtin_bit_cast(int, h2rcp(__builtin_bit_cast(half2, a)));
}
template <>
__quickreduce_device_inline__ int packed_rcp<nv_bfloat16>(int a) {
bf162_int_union A, R;
A.i = a;
R.bf2 = h2rcp(A.bf2);
return R.i;
}
// changes dtype
__quickreduce_device_inline__ float T2float_cast(half a) {
return __half2float(a);
}
__quickreduce_device_inline__ float T2float_cast(nv_bfloat16 a) {
return __bfloat162float(a);
}
template <typename T>
__quickreduce_device_inline__ int group_abs_max(int32x4_t atom) {
const int group_leader = (threadIdx.x / kThreadGroupSize) * kThreadGroupSize;
int wmax, wmin, wblockmax;
int a, b;
a = packed_max<T>(atom[0], atom[1]);
b = packed_max<T>(atom[2], atom[3]);
wmax = packed_max<T>(a, b);
a = packed_min<T>(atom[0], atom[1]);
b = packed_min<T>(atom[2], atom[3]);
wmin = packed_min<T>(a, b);
// Reduce the max among a group of threads
// Note: This is basically 2 blocks of values setup as the
// upper/lower halves of the f16x2_t
for (int i = 1; i < kThreadGroupSize; i <<= 1) {
int x = __shfl_down(wmax, i);
wmax = packed_max<T>(wmax, x);
int y = __shfl_down(wmin, i);
wmin = packed_min<T>(wmin, y);
}
wblockmax = packed_abs_max<T>(wmax, wmin);
// Share with the cohort
wblockmax = __shfl(wblockmax, group_leader);
return wblockmax;
}
__quickreduce_device_inline__ void set_sync_flag(uint32_t* flag_ptr,
uint32_t flag) {
__atomic_store_n(flag_ptr, flag, __ATOMIC_RELEASE);
}
__quickreduce_device_inline__ void wait_sync_flag(uint32_t* flag_ptr,
uint32_t flag) {
while (__atomic_load_n(flag_ptr, __ATOMIC_RELAXED) != flag) {
}
}
} // namespace quickreduce
\ No newline at end of file
csrc/quickreduce/quick_reduce.h 0 → 100644
View file @ 99324e25
#pragma once
#include <vector>
#include <hip/hip_runtime.h>
#include "quick_reduce_impl.cuh"
#define HIP_CHECK(err) \
do { \
hipError_t err_ = (err); \
if (err_ != hipSuccess) { \
std::printf("HIP error %d at %s:%d. %s\n", err_, __FILE__, __LINE__, \
hipGetErrorString(err_)); \
throw std::runtime_error("HIP error"); \
} \
} while (0)
namespace quickreduce {
using fptr_t = int64_t;
static_assert(sizeof(void*) == sizeof(fptr_t));
template <typename AllReduceKernel, typename T>
__global__ __quickreduce_launch_bounds_two_shot__ static void
allreduce_prototype_twoshot(T const* A, T* B, uint32_t N, uint32_t num_blocks,
int rank, uint8_t** dbuffer_list,
uint32_t data_offset, uint32_t flag_color) {
int block = blockIdx.x;
int grid = gridDim.x;
while (block < num_blocks) {
AllReduceKernel::run(A, B, N, block, rank, dbuffer_list, data_offset,
flag_color);
block += grid;
flag_color++;
}
}
#define TWOSHOT_DISPATCH(__codec) \
if (world_size == 2) { \
using LineCodec = __codec<T, 2>; \
using AllReduceKernel = AllReduceTwoshot<T, LineCodec, cast_bf2half>; \
hipLaunchKernelGGL((allreduce_prototype_twoshot<AllReduceKernel, T>), \
dim3(grid), dim3(kBlockTwoShot), 0, stream, A, B, N, \
num_blocks, rank, dbuffer_list, data_offset, \
flag_color); \
} else if (world_size == 4) { \
using LineCodec = __codec<T, 4>; \
using AllReduceKernel = AllReduceTwoshot<T, LineCodec, cast_bf2half>; \
hipLaunchKernelGGL((allreduce_prototype_twoshot<AllReduceKernel, T>), \
dim3(grid), dim3(kBlockTwoShot), 0, stream, A, B, N, \
num_blocks, rank, dbuffer_list, data_offset, \
flag_color); \
} else if (world_size == 8) { \
using LineCodec = __codec<T, 8>; \
using AllReduceKernel = AllReduceTwoshot<T, LineCodec, cast_bf2half>; \
hipLaunchKernelGGL((allreduce_prototype_twoshot<AllReduceKernel, T>), \
dim3(grid), dim3(kBlockTwoShot), 0, stream, A, B, N, \
num_blocks, rank, dbuffer_list, data_offset, \
flag_color); \
}
enum QuickReduceQuantLevel {
F16 = 0,
INT8 = 1,
INT6 = 2,
INT4 = 3,
};
struct DeviceComms {
// Max problem size is 2GB (in bytes) or half of uint32_t max value.
int64_t kMaxProblemSize =
static_cast<int64_t>(std::numeric_limits<int32_t>::max()) + 1;
// Max TP-8
static int constexpr kMaxWorldSize = 8;
bool initialized = false;
uint32_t flag_color = 1;
int world_size;
int rank;
uint8_t* dbuffer;
uint8_t** dbuffer_list;
hipIpcMemHandle_t buffer_ipc_handle;
std::vector<hipIpcMemHandle_t> all_buffer_ipc_handles;
std::vector<uint8_t*> buffer_list;
uint32_t data_offset;
DeviceComms() : initialized(false), world_size(1), rank(0) {}
~DeviceComms() { destroy(); }
void init(int world_size, int rank,
std::optional<int64_t> max_problem_size = std::nullopt) {
destroy();
this->world_size = world_size;
this->rank = rank;
if (max_problem_size.has_value() && max_problem_size.value() > 0) {
this->kMaxProblemSize = max_problem_size.value();
}
// Allocate buffer size for worst case: F16 2-stage buffer.
uint32_t flags_buffer_size =
2 * world_size * kMaxNumBlocks * sizeof(uint32_t);
static int64_t data_buffer_size = 2 * this->kMaxProblemSize;
int64_t total_buffer_size = flags_buffer_size + data_buffer_size;
data_offset = flags_buffer_size;
HIP_CHECK(hipExtMallocWithFlags((void**)&dbuffer, total_buffer_size,
hipDeviceMallocUncached));
// Clear the flags buffer.
HIP_CHECK(hipMemset(dbuffer, 0, flags_buffer_size));
// Device-side list of IPC buffers.
buffer_list.resize(world_size);
HIP_CHECK(hipMalloc(&dbuffer_list, world_size * sizeof(uint8_t*)));
// Create IPC handles for rank's communication buffer.
all_buffer_ipc_handles.resize(world_size);
HIP_CHECK(hipIpcGetMemHandle(&buffer_ipc_handle, dbuffer));
initialized = true;
}
int get_world_size() { return world_size; }
int get_rank() { return rank; }
bool status() { return initialized; }
hipIpcMemHandle_t const get_handle() { return buffer_ipc_handle; }
void destroy() {
if (initialized) {
for (int i = 0; i < world_size; i++) {
if (i != rank) {
HIP_CHECK(hipIpcCloseMemHandle(dbuffer_list[i]));
}
}
HIP_CHECK(hipFree(dbuffer));
HIP_CHECK(hipFree(dbuffer_list));
initialized = false;
}
}
void open_ipc_handles(std::vector<hipIpcMemHandle_t> const& ipc_handles) {
assert(ipc_handles.size() == all_buffer_ipc_handles.size());
for (int i = 0; i < world_size; i++) {
all_buffer_ipc_handles[i] = ipc_handles[i];
}
// Open device memory access to the IPC communication buffers.
// Note: For our own rank, we do not need to open a handle.
for (int i = 0; i < world_size; i++) {
if (i != rank) {
HIP_CHECK(hipIpcOpenMemHandle((void**)&buffer_list[i],
all_buffer_ipc_handles[i],
hipIpcMemLazyEnablePeerAccess));
} else {
buffer_list[i] = dbuffer;
}
}
HIP_CHECK(hipMemcpy(dbuffer_list, buffer_list.data(),
world_size * sizeof(uint8_t*), hipMemcpyHostToDevice));
}
template <typename T, bool cast_bf2half>
void allreduce(T const* A, T* B, uint32_t N, int quant_level,
hipStream_t stream) {
if (world_size != 2 && world_size != 4 && world_size != 8) {
throw std::runtime_error("All Reduce not supported for world_size = " +
std::to_string(world_size));
}
// Configuration.
uint32_t msg_size = N * sizeof(T);
uint32_t num_blocks = divceil(msg_size, kTileSize);
uint32_t grid = min(kMaxNumBlocks, num_blocks);
auto quant_level_ = static_cast<QuickReduceQuantLevel>(quant_level);
switch (quant_level_) {
case QuickReduceQuantLevel::INT8:
TWOSHOT_DISPATCH(CodecQ8)
break;
case QuickReduceQuantLevel::INT6:
TWOSHOT_DISPATCH(CodecQ6)
break;
case QuickReduceQuantLevel::INT4:
TWOSHOT_DISPATCH(CodecQ4)
break;
default:
TWOSHOT_DISPATCH(CodecFP)
break;
}
HIP_CHECK(cudaGetLastError());
// Rotate the flag color.
flag_color += divceil(N, grid);
}
};
} // namespace quickreduce
\ No newline at end of file
csrc/quickreduce/quick_reduce_impl.cuh 0 → 100644
View file @ 99324e25
#pragma once
#include <hip/hip_runtime.h>
#include "base.h"
namespace quickreduce {
struct CodecBase {
const int thread;
const int rank;
const int group_leader;
__quickreduce_device_inline__ CodecBase(int thread, int rank)
: thread(thread),
rank(rank),
group_leader((threadIdx.x / kThreadGroupSize) * kThreadGroupSize) {
set_fp16_ovfl(true);
}
};
// Default full precision codec.
template <typename T, int world_size>
struct CodecFP : public CodecBase {
static constexpr int kWorldSize = world_size;
static constexpr int kRankAtoms = kAtoms / kWorldSize;
// Codec tile size process by this workgroup.
// Each thread processes atoms of f16x8_t (16B).
static constexpr int kRankTransmittedTileSize =
kBlockSize * kRankAtoms * sizeof(int32x4_t);
static_assert(kRankTransmittedTileSize % 16 == 0,
"kRankTransmittedTileSize must be 16B aligned.");
// Total tile size for the collective communication.
static constexpr int kTransmittedTileSize =
kRankTransmittedTileSize * kWorldSize;
__quickreduce_device_inline__ CodecFP(int thread, int rank)
: CodecBase(thread, rank) {}
__quickreduce_device_inline__ void send(int32x4_t* __restrict__ send_buffer,
const int32x4_t* __restrict__ data) {
for (int i = 0; i < kRankAtoms; i++) {
__builtin_nontemporal_store(data[i], send_buffer + thread);
send_buffer += kAtomStride;
}
}
__quickreduce_device_inline__ void recv(int32x4_t** __restrict__ recv_buffer,
int32x4_t* __restrict__ data) {
for (int i = 0; i < kRankAtoms; i++) {
data[i] = __builtin_nontemporal_load(*recv_buffer + thread);
*recv_buffer += kAtomStride;
}
}
};
// Int4 symmetric quantization codec.
// We quantize the FP16 data to block-scaled Int4 in blocks of 4 *
// kThreadGroupSize.
template <typename T, int world_size>
struct CodecQ4 : public CodecBase {
static constexpr int kWorldSize = world_size;
// Codec tile size process by this workgroup.
// Each threads processes a fragment of fp16x8_t (16B),
// into a int4x8_t (4B) and a fp16 scale shared among 32 values.
static constexpr int kRankAtoms = kAtoms / kWorldSize;
static constexpr int kRankTileStride = 1152;
static constexpr int kRankTileScaleOffset = 1024;
static constexpr int kRankTransmittedTileSize = kRankTileStride * kRankAtoms;
static_assert(kRankTransmittedTileSize % 16 == 0,
"kRankTransmittedTileSize must be 16B aligned.");
static constexpr int kRankBufferTileStride =
kRankTileStride / sizeof(int32x4_t);
// Total tile size for the collective communication.
static constexpr int kTransmittedTileSize =
kRankTransmittedTileSize * kWorldSize;
// Constants configuration
// {-1/8.0h, -1/8.0h}, f16x2_t
static constexpr int kScaleFactor =
std::is_same<T, half>::value ? 0xB000B000 : 0xBE00BE00;
// {1e-7, 1e-7}, f16x2_t
static constexpr int kScaleEpsilon =
std::is_same<T, half>::value ? 0x00010001 : 0x33D733D7;
// {-8, -8}, f16x2_t
static constexpr int kRangeMin =
std::is_same<T, half>::value ? 0xC800C800 : 0xC100C100;
// {+7, +7}, f16x2_t
static constexpr int kRangeMax =
std::is_same<T, half>::value ? 0x47004700 : 0x40E040E0;
// {+8, +8}, int16x2_t
static constexpr int kRangeBias = 0x00080008;
__quickreduce_device_inline__ CodecQ4(int thread, int rank)
: CodecBase(thread, rank) {}
__quickreduce_device_inline__ void send(int32x4_t* __restrict__ send_buffer,
const int32x4_t* __restrict__ data) {
for (int k = 0; k < kRankAtoms; k++) {
int32x4_t const atom = data[k];
// Compute the absolute maximum of the atom in the thread group
// In 2 blocks of values, upper/lower halves of the f16x2_t
int wblockmax = group_abs_max<T>(atom);
// Derive scales
int decoding_scale;
int encoding_scale;
decoding_scale = packed_mul<T>(wblockmax, kScaleFactor);
encoding_scale = packed_add<T>(decoding_scale, kScaleEpsilon);
encoding_scale = packed_rcp<T>(encoding_scale);
// Apply scales to get quantized values
int32x4_t w;
for (int i = 0; i < 4; i++) {
w[i] = packed_mul<T>(atom[i], encoding_scale);
w[i] = packed_max<T>(w[i], kRangeMin);
w[i] = packed_min<T>(w[i], kRangeMax);
}
// Convert from f16x2_t to uint16x2_t
int32x4_t q;
{
int16_t* qi = reinterpret_cast<int16_t*>(&q);
T* wh = reinterpret_cast<T*>(&w);
for (int i = 0; i < 8; i++) qi[i] = (int16_t)rintf(T2float_cast(wh[i]));
for (int i = 0; i < 4; i++) {
q[i] = packed_add<int16_t>(q[i], kRangeBias);
}
}
// Pack 8 x q4 into int32_t
int qw = q[0] | (q[1] << 4) | (q[2] << 8) | (q[3] << 12);
// Write quantized atom to send_buffer
// note: only the group leader stores the scale
uint8_t* atom_ptr =
reinterpret_cast<uint8_t*>(send_buffer + k * kRankBufferTileStride);
int32_t* qw_ptr = reinterpret_cast<int32_t*>(atom_ptr) + thread;
int* qs_ptr = reinterpret_cast<int*>(atom_ptr + kRankTileScaleOffset) +
(thread / 8);
__builtin_nontemporal_store(qw, qw_ptr);
if (threadIdx.x == group_leader) {
__builtin_nontemporal_store(decoding_scale, qs_ptr);
}
}
}
__quickreduce_device_inline__ void recv(int32x4_t** __restrict__ recv_buffer,
int32x4_t* __restrict__ data) {
for (int k = 0; k < kRankAtoms; k++) {
// Directly read quantized atom from recv_buffer
uint8_t* atom_ptr = reinterpret_cast<uint8_t*>(*recv_buffer);
int32_t* qw_ptr = reinterpret_cast<int32_t*>(atom_ptr) + thread;
int* qs_ptr = reinterpret_cast<int*>(atom_ptr + kRankTileScaleOffset) +
(thread / 8);
int32_t qw = __builtin_nontemporal_load(qw_ptr);
int qs = __builtin_nontemporal_load(qs_ptr);
*recv_buffer += kRankBufferTileStride;
// Unpack q4 into f16x8_t
int32x4_t w;
{
static constexpr uint kMask000F = 0x000F000F;
static constexpr uint kHalf2_1024 =
0x64006400; // {1024.0, 1024.0}, fp16x2_t
static uint constexpr kHalf2_1032 =
0xE408E408; // {-1032.0, -1032.0}, fp16x2_t
for (int i = 0; i < 4; i++) {
if constexpr (std::is_same<T, half>::value) {
int32_t q4 = ((qw >> (i * 4)) & kMask000F) | kHalf2_1024;
w[i] = packed_add<half>(q4, kHalf2_1032);
} else {
int32_t int16_2 = (qw >> (i * 4)) & kMask000F;
int16_t low = static_cast<int16_t>(int16_2 & 0xFFFF);
int16_t high = static_cast<int16_t>((int16_2 >> 16) & 0xFFFF);
nv_bfloat16 bf_low = __float2bfloat16(static_cast<float>(low));
nv_bfloat16 bf_high = __float2bfloat16(static_cast<float>(high));
nv_bfloat162 bf2 = __halves2bfloat162(bf_low, bf_high);
int32_t packed_bf16 = *reinterpret_cast<int32_t*>(&bf2);
w[i] = packed_add<nv_bfloat16>(packed_bf16, kRangeMin);
}
}
}
// Apply decoding scales
for (int i = 0; i < 4; i++) {
w[i] = packed_mul<T>(w[i], qs);
}
data[k] = w;
}
}
};
// Int6 symmetric quantization codec.
// We quantize the FP16 data to block-scaled Int6 in blocks of 4 *
// kThreadGroupSize.
template <typename T, int world_size>
struct CodecQ6 : public CodecBase {
static constexpr int kWorldSize = world_size;
// Codec tile size process by this workgroup.
// Each threads processes a fragment of fp16x8_t (16B),
// into a int6x8_t (4B + 2B) and a fp16 scale shared among 32 values.
static constexpr int kRankAtoms = kAtoms / kWorldSize;
static constexpr int kRankTileStride = 1664;
static constexpr int kRankTileQ2Offset = 1024;
static constexpr int kRankTileScaleOffset = 1536;
static constexpr int kRankTransmittedTileSize = kRankTileStride * kRankAtoms;
static_assert(kRankTransmittedTileSize % 16 == 0,
"kRankTransmittedTileSize must be 16B aligned.");
static constexpr int kRankBufferTileStride =
kRankTileStride / sizeof(int32x4_t);
// Total tile size for the collective communication.
static constexpr int kTransmittedTileSize =
kRankTransmittedTileSize * kWorldSize;
// Constants configuration
// {-1/32.0h, -1/32.0h}, fp16x2_t
static constexpr int kScaleFactor =
std::is_same<T, half>::value ? 0xA800A800 : 0xBD00BD00;
// {1e-7, 1e-7}, fp16x2_t
static constexpr int kScaleEpsilon =
std::is_same<T, half>::value ? 0x00010001 : 0x33D733D7;
// {-32, -32}, fp16x2_t
static constexpr int kRangeMin =
std::is_same<T, half>::value ? 0xD000D000 : 0xC200C200;
// {+31, +31}, fp16x2_t
static constexpr int kRangeMax =
std::is_same<T, half>::value ? 0x4FC04FC0 : 0x41F841F8;
// {+32, +32}, int16x2_t
static constexpr int kRangeBias = 0x00200020;
__quickreduce_device_inline__ CodecQ6(int thread, int rank)
: CodecBase(thread, rank) {}
__quickreduce_device_inline__ void send(int32x4_t* __restrict__ send_buffer,
const int32x4_t* __restrict__ data) {
for (int k = 0; k < kRankAtoms; k++) {
int32x4_t const atom = data[k];
// Compute the absolute maximum of the atom in the thread group
// In 2 blocks of values, upper/lower halves of the f16x2_t
int wblockmax = group_abs_max<T>(atom);
// Derive scales
int decoding_scale;
int encoding_scale;
decoding_scale = packed_mul<T>(wblockmax, kScaleFactor);
encoding_scale = packed_add<T>(decoding_scale, kScaleEpsilon);
encoding_scale = packed_rcp<T>(encoding_scale);
// Apply scales to get quantized values
int32x4_t w;
for (int i = 0; i < 4; i++) {
w[i] = packed_mul<T>(atom[i], encoding_scale);
w[i] = packed_max<T>(w[i], kRangeMin);
w[i] = packed_min<T>(w[i], kRangeMax);
}
// Convert from f16x2_t to uint16x2_t
int32x4_t q;
{
int16_t* qi = reinterpret_cast<int16_t*>(&q);
T* wh = reinterpret_cast<T*>(&w);
for (int i = 0; i < 8; i++) qi[i] = (int16_t)rintf(T2float_cast(wh[i]));
for (int i = 0; i < 4; i++) {
q[i] = packed_add<int16_t>(q[i], kRangeBias);
}
}
// Pack 8 x q6 into int32_t + int16_t
uint32_t q4w;
uint16_t q2w = 0;
q4w = (q[0] & 0x000F000F) | ((q[1] & 0x000F000F) << 4) |
((q[2] & 0x000F000F) << 8) | ((q[3] & 0x000F000F) << 12);
{
int16_t* tw = reinterpret_cast<int16_t*>(&q);
#pragma unroll
for (int i = 0; i < 8; i++) {
q2w |= (tw[i] >> 4) << (i * 2);
}
}
// Write quantized atom to send_buffer
// note: only the group leader stores the scale
uint8_t* atom_ptr =
reinterpret_cast<uint8_t*>(send_buffer + k * kRankBufferTileStride);
uint32_t* q4w_ptr = reinterpret_cast<uint32_t*>(atom_ptr) + thread;
uint16_t* q2w_ptr =
reinterpret_cast<uint16_t*>(atom_ptr + kRankTileQ2Offset) + thread;
int* qs_ptr = reinterpret_cast<int*>(atom_ptr + kRankTileScaleOffset) +
(thread / 8);
__builtin_nontemporal_store(q4w, q4w_ptr);
__builtin_nontemporal_store(q2w, q2w_ptr);
if (threadIdx.x == group_leader) {
__builtin_nontemporal_store(decoding_scale, qs_ptr);
}
}
}
__quickreduce_device_inline__ void recv(int32x4_t** __restrict__ recv_buffer,
int32x4_t* __restrict__ data) {
for (int k = 0; k < kRankAtoms; k++) {
// Directly read quantized atom from recv_buffer
uint8_t* atom_ptr = reinterpret_cast<uint8_t*>(*recv_buffer);
uint32_t* q4w_ptr = reinterpret_cast<uint32_t*>(atom_ptr) + thread;
uint16_t* q2w_ptr =
reinterpret_cast<uint16_t*>(atom_ptr + kRankTileQ2Offset) + thread;
int* qs_ptr = reinterpret_cast<int*>(atom_ptr + kRankTileScaleOffset) +
(thread / 8);
uint32_t q4w = __builtin_nontemporal_load(q4w_ptr);
uint16_t q2w = __builtin_nontemporal_load(q2w_ptr);
int qs = __builtin_nontemporal_load(qs_ptr);
*recv_buffer += kRankBufferTileStride;
// Unpack q6 into fp16x8_t
int32x4_t w;
{
static uint constexpr kMask000F = 0x000F000F;
static uint constexpr kHalf2_1024 =
0x64006400; // {1024.0, 1024.0}, fp16x2_t
static uint constexpr kHalf2_1056 =
0xE420E420; // {-1056.0, -1056.0}, fp16x2_t
#pragma unroll
for (int i = 0; i < 4; i++) {
int32_t q4 = q4w & kMask000F;
int32_t q2 = (q2w & 0x3) | ((q2w & 0xC) << 14);
q4w >>= 4;
q2w >>= 4;
if constexpr (std::is_same<T, half>::value) {
int32_t q6 = q4 | (q2 << 4) | kHalf2_1024;
asm volatile("v_pk_add_f16 %0, %1, %2"
: "=v"(w[i])
: "v"(q6), "v"(kHalf2_1056));
} else {
int32_t int16_2 = q4 | (q2 << 4);
int16_t low = static_cast<int16_t>(int16_2 & 0xFFFF);
int16_t high = static_cast<int16_t>((int16_2 >> 16) & 0xFFFF);
nv_bfloat16 bf_low = __float2bfloat16(static_cast<float>(low));
nv_bfloat16 bf_high = __float2bfloat16(static_cast<float>(high));
nv_bfloat162 bf2 = __halves2bfloat162(bf_low, bf_high);
int32_t packed_bf16 = *reinterpret_cast<int32_t*>(&bf2);
w[i] = packed_add<nv_bfloat16>(packed_bf16, kRangeMin);
}
}
}
// Apply decoding scales
for (int i = 0; i < 4; i++) {
w[i] = packed_mul<T>(w[i], qs);
}
// That's pretty much it...
data[k] = w;
}
}
};
// Int8 symmetric quantization codec.
// We quantize the FP16 data to block-scaled Int8 in blocks of 4 *
// kThreadGroupSize.
template <typename T, int world_size>
struct CodecQ8 : public CodecBase {
static constexpr int kWorldSize = world_size;
// Codec tile size process by this workgroup.
// Each threads processes a fragment of f16x8_t (16B),
// into a int8x8_t (8B) and a f16 scale shared among 32 values.
static constexpr int kRankAtoms = kAtoms / kWorldSize;
static constexpr int kRankTileStride = 2176;
static constexpr int kRankTileScaleOffset = 2048;
static constexpr int kRankTransmittedTileSize = kRankTileStride * kRankAtoms;
static_assert(kRankTransmittedTileSize % 16 == 0,
"kRankTileSize must be 16B aligned.");
static constexpr int kRankBufferTileStride =
kRankTileStride / sizeof(int32x4_t);
// Total tile size for the collective communication.
static constexpr int kTransmittedTileSize =
kRankTransmittedTileSize * kWorldSize;
// Constants configuration
// {-1/128.0h, -1/128.0h}, f16x2_t
static constexpr int kScaleFactor =
std::is_same<T, half>::value ? 0xA000A000 : 0xBC00BC00;
// {1e-7, 1e-7}, f16x2_t
static constexpr int kScaleEpsilon =
std::is_same<T, half>::value ? 0x00010001 : 0x33D733D7;
// {-128, -128}, f16x2_t
static constexpr int kRangeMin =
std::is_same<T, half>::value ? 0xD800D800 : 0xC300C300;
// {+127, +127}, f16x2_t
static constexpr int kRangeMax =
std::is_same<T, half>::value ? 0x57F057F0 : 0x42FE42FE;
// {+128, +128}, int16x2_t
static constexpr int kRangeBias = 0x00800080;
__quickreduce_device_inline__ CodecQ8(int thread, int rank)
: CodecBase(thread, rank) {}
__quickreduce_device_inline__ void send(int32x4_t* __restrict__ send_buffer,
int32x4_t const* __restrict__ data) {
for (int k = 0; k < kRankAtoms; k++) {
int32x4_t const atom = data[k];
// Compute the absolute maximum of the atom in the thread group
// In 2 blocks of values, upper/lower halves of the f16x2_t
int wblockmax = group_abs_max<T>(atom);
// Derive scales
int decoding_scale;
int encoding_scale;
decoding_scale = packed_mul<T>(wblockmax, kScaleFactor);
encoding_scale = packed_add<T>(decoding_scale, kScaleEpsilon);
encoding_scale = packed_rcp<T>(encoding_scale);
// Apply scales to get quantized values
int32x4_t w;
for (int i = 0; i < 4; i++) {
w[i] = packed_mul<T>(atom[i], encoding_scale);
w[i] = packed_max<T>(w[i], kRangeMin);
w[i] = packed_min<T>(w[i], kRangeMax);
}
// Convert from f16x2_t to uint16x2_t
int32x4_t q;
{
int16_t* qi = reinterpret_cast<int16_t*>(&q);
T* wh = reinterpret_cast<T*>(&w);
for (int i = 0; i < 8; i++) qi[i] = (int16_t)rintf(T2float_cast(wh[i]));
for (int i = 0; i < 4; i++) {
q[i] = packed_add<int16_t>(q[i], kRangeBias);
}
}
// Pack 8 x q8 into int32x2_t
int32x2_t qw;
qw[0] = q[0] | (q[1] << 8);
qw[1] = q[2] | (q[3] << 8);
// Write quantized atom to send_buffer
// note: only the group leader stores the scale
uint8_t* atom_ptr =
reinterpret_cast<uint8_t*>(send_buffer + k * kRankBufferTileStride);
int32x2_t* qw_ptr = reinterpret_cast<int32x2_t*>(atom_ptr) + thread;
int* qs_ptr = reinterpret_cast<int*>(atom_ptr + kRankTileScaleOffset) +
(thread / 8);
__builtin_nontemporal_store(qw, qw_ptr);
if (threadIdx.x == group_leader) {
__builtin_nontemporal_store(decoding_scale, qs_ptr);
}
}
}
__quickreduce_device_inline__ void recv(int32x4_t** __restrict__ recv_buffer,
int32x4_t* __restrict__ data) {
for (int k = 0; k < kRankAtoms; k++) {
// Directly read quantized atom from recv_buffer
uint8_t* atom_ptr = reinterpret_cast<uint8_t*>(*recv_buffer);
int32x2_t* qw_ptr = reinterpret_cast<int32x2_t*>(atom_ptr) + thread;
int* qs_ptr = reinterpret_cast<int*>(atom_ptr + kRankTileScaleOffset) +
(thread / 8);
int32x2_t qw = __builtin_nontemporal_load(qw_ptr);
int qs = __builtin_nontemporal_load(qs_ptr);
*recv_buffer += kRankBufferTileStride;
// Unpack q8 into fp16x8_t
int32x4_t w;
{
static uint constexpr kMask00FF = 0x00FF00FF;
// {1024.0, 1024.0}, fp16x2_t
static uint constexpr kHalf2_1024 = 0x64006400;
// {-1152.0, -1152.0}, fp16x2_t
static uint constexpr kHalf2_1152 = 0xE480E480;
#pragma unroll
for (int i = 0; i < 4; i++) {
if constexpr (std::is_same<T, half>::value) {
int32_t q8 =
((qw[i / 2] >> ((i % 2) * 8)) & kMask00FF) | kHalf2_1024;
w[i] = packed_add<half>(q8, kHalf2_1152);
} else {
int32_t int16_2 = (qw[i / 2] >> ((i % 2) * 8)) & kMask00FF;
int16_t low = static_cast<int16_t>(int16_2 & 0xFFFF);
int16_t high = static_cast<int16_t>((int16_2 >> 16) & 0xFFFF);
nv_bfloat16 bf_low = __float2bfloat16(static_cast<float>(low));
nv_bfloat16 bf_high = __float2bfloat16(static_cast<float>(high));
nv_bfloat162 bf2 = __halves2bfloat162(bf_low, bf_high);
int32_t packed_bf16 = *reinterpret_cast<int32_t*>(&bf2);
w[i] = packed_add<nv_bfloat16>(packed_bf16, kRangeMin);
}
}
}
// Apply decoding scales
for (int i = 0; i < 4; i++) {
w[i] = packed_mul<T>(w[i], qs);
}
data[k] = w;
}
}
};
// Twoshot All Reduce
template <typename T, class Codec, bool cast_bf2half>
struct AllReduceTwoshot {
static_assert(sizeof(T) == 2);
static constexpr int kWorldSize = Codec::kWorldSize;
__device__ static void run(
T const* __restrict__ input, T* __restrict__ output,
uint32_t const N, // number of elements
int const block, // block index
int const rank, // rank index
uint8_t** __restrict__ buffer_list, // communication buffers
uint32_t const data_offset, // offset to start of the data buffer
uint32_t flag_color) {
// Topology
int thread = threadIdx.x + threadIdx.y * kWavefront;
uint8_t* rank_buffer = buffer_list[rank];
Codec codec(thread, rank);
int block_id = blockIdx.x;
int grid_size = gridDim.x;
// --------------------------------------------------------
// Read input into registers
int32x4_t tA[kAtoms];
BufferResource src_buffer(const_cast<T*>(input), N * sizeof(T));
uint32_t src_offset = block * kTileSize + thread * sizeof(int32x4_t);
for (int i = 0; i < kAtoms; i++) {
tA[i] = buffer_load_dwordx4(src_buffer.descriptor, src_offset, 0, 0);
src_offset += kAtomStride * sizeof(int32x4_t);
if constexpr (cast_bf2half) {
const nv_bfloat162* bf_buf =
reinterpret_cast<const nv_bfloat162*>(&tA[i]);
half2 half_buf[4];
#pragma unroll
for (int j = 0; j < 4; ++j) {
float2 f = __bfloat1622float2(bf_buf[j]);
half_buf[j] = __float22half2_rn(f);
}
tA[i] = *reinterpret_cast<const int32x4_t*>(half_buf);
}
}
// --------------------------------------------------------
// Phase-1A: Write segment data into the communication buffer of the target
// rank responsible for this segment.
uint32_t comm_data0_offset =
data_offset + block_id * Codec::kTransmittedTileSize;
uint32_t comm_data1_offset =
grid_size * Codec::kTransmittedTileSize + comm_data0_offset;
uint32_t comm_flags0_offset = block_id * (kWorldSize * sizeof(uint32_t));
uint32_t comm_flags1_offset =
grid_size * (kWorldSize * sizeof(uint32_t)) + comm_flags0_offset;
for (int r = 0; r < kWorldSize; r++) {
int32x4_t* send_buffer =
reinterpret_cast<int32x4_t*>(buffer_list[r] + comm_data0_offset +
rank * Codec::kRankTransmittedTileSize);
codec.send(send_buffer, &tA[r * Codec::kRankAtoms]);
}
__syncthreads();
if (thread < kWorldSize) {
int r = thread;
uint32_t* flag_ptr = reinterpret_cast<uint32_t*>(
buffer_list[r] + comm_flags0_offset + rank * sizeof(uint32_t));
set_sync_flag(flag_ptr, flag_color);
}
// --------------------------------------------------------
// Phase-1B: Reduce the segment data from the communication buffers.
int32x4_t tR[Codec::kRankAtoms] = {};
{
// Read the data from the communication buffer.
int32x4_t* recv_buffer =
reinterpret_cast<int32x4_t*>(rank_buffer + comm_data0_offset);
uint32_t* flag_ptr =
reinterpret_cast<uint32_t*>(rank_buffer + comm_flags0_offset);
for (int r = 0; r < kWorldSize; r++) {
// Wait for the flags to be set.
if (thread == 0) {
wait_sync_flag(&flag_ptr[r], flag_color);
}
__syncthreads();
// note: we reuse tA as temp buffer here
codec.recv(&recv_buffer, tA);
for (int i = 0; i < Codec::kRankAtoms; i++) {
packed_assign_add<T>(&tR[i], &tA[i]);
}
}
}
// Phase-2: Write the reduced segment to every other rank
for (int r = 0; r < kWorldSize; r++) {
int32x4_t* send_buffer =
reinterpret_cast<int32x4_t*>(buffer_list[r] + comm_data1_offset +
rank * Codec::kRankTransmittedTileSize);
codec.send(send_buffer, tR);
}
__syncthreads();
if (thread < kWorldSize) {
int r = thread;
uint32_t* flag_ptr = reinterpret_cast<uint32_t*>(
buffer_list[r] + comm_flags1_offset + rank * sizeof(uint32_t));
set_sync_flag(flag_ptr, flag_color);
}
// Phase-2: Read the gather segments from the rank's communication buffer.
{
// Read the data from the communication buffer.
int32x4_t* recv_buffer =
reinterpret_cast<int32x4_t*>(rank_buffer + comm_data1_offset);
uint32_t* flag_ptr =
reinterpret_cast<uint32_t*>(rank_buffer + comm_flags1_offset);
for (int r = 0; r < kWorldSize; r++) {
// Wait for the flags to be set.
if (thread == 0) {
wait_sync_flag(&flag_ptr[r], flag_color);
}
__syncthreads();
// Gather all reduced and final rank segments into tA.
codec.recv(&recv_buffer, &tA[r * Codec::kRankAtoms]);
}
}
// --------------------------------------------------------
// Write the result to output.
BufferResource dst_buffer(output, N * sizeof(T));
uint32_t dst_offset = block * kTileSize + thread * sizeof(int32x4_t);
for (int i = 0; i < kAtoms; i++) {
if constexpr (cast_bf2half) {
const half2* half_buf = reinterpret_cast<const half2*>(&tA[i]);
nv_bfloat162 bf16_buf[4];
#pragma unroll
for (int j = 0; j < 4; ++j) {
float2 f = __half22float2(half_buf[j]);
bf16_buf[j] = __float22bfloat162_rn(f);
}
buffer_store_dwordx4(*reinterpret_cast<const int32x4_t*>(bf16_buf),
dst_buffer.descriptor, dst_offset, 0, 0);
} else {
buffer_store_dwordx4(tA[i], dst_buffer.descriptor, dst_offset, 0, 0);
}
dst_offset += kAtomStride * sizeof(int32x4_t);
}
}
};
} // namespace quickreduce
\ No newline at end of file
csrc/rocm/attention.cu
View file @ 99324e25
......@@ -136,11 +136,6 @@ __device__ __forceinline__ T from_float(const float& inp) {
template <typename T>
__device__ __forceinline__ _B16x4 from_floatx4(const floatx4& inp) {
[[maybe_unused]] union tmpcvt {
uint16_t u;
_Float16 f;
__hip_bfloat16 b;
} t16;
_B16x4 ret;
if constexpr (std::is_same<T, _Float16>::value) {
union h2cvt {
......@@ -169,11 +164,6 @@ __device__ __forceinline__ _B16x4 from_floatx4(const floatx4& inp) {
template <typename T>
__device__ __forceinline__ _B16x4 addx4(const _B16x4& inp1,
const _B16x4& inp2) {
[[maybe_unused]] union tmpcvt {
uint16_t u;
_Float16 f;
__hip_bfloat16 b;
} t1, t2, res;
_B16x4 ret;
if constexpr (std::is_same<T, _Float16>::value) {
union h2cvt {
......@@ -325,8 +315,6 @@ __launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel(
constexpr int GQA_RATIO4 = DIVIDE_ROUND_UP(GQA_RATIO, 4);
[[maybe_unused]] __shared__ float shared_qk_max[NWARPS][16 + 1];
[[maybe_unused]] __shared__ float shared_exp_sum[NWARPS][16 + 1];
// shared_logits is used for multiple purposes
__shared__ _B16x4 shared_logits[NWARPS][4][16][4];
......@@ -444,8 +432,6 @@ __launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel(
const cache_t* k_ptr2 = k_ptr + kblock_number * kv_block_stride;
const int klocal_token_idx =
TOKENS_PER_WARP * warpid + token_depth * 16 + lane16id;
[[maybe_unused]] const int kglobal_token_idx =
partition_start_token_idx + klocal_token_idx;
const int kphysical_block_offset = klocal_token_idx % BLOCK_SIZE;
const cache_t* k_ptr3 = k_ptr2 + kphysical_block_offset * KX;
......@@ -1309,9 +1295,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
const int context_len = context_lens[seq_idx];
const int num_partitions = DIVIDE_ROUND_UP(context_len, PARTITION_SIZE);
[[maybe_unused]] constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
const auto warpid = threadIdx.x / WARP_SIZE;
[[maybe_unused]] const auto laneid = threadIdx.x % WARP_SIZE;
__shared__ float shared_global_exp_sum;
// max num partitions supported is warp_size * NPAR_LOOPS
......@@ -1614,7 +1598,6 @@ __launch_bounds__(NUM_THREADS, 3) void paged_attention_ll4mi_QKV_mfma16_kernel(
const int warpid = threadIdx.x / WARP_SIZE;
const int laneid = threadIdx.x % WARP_SIZE;
const int lane2id = laneid % 2;
const int lane4id = laneid % 4;
const int lane16id = laneid % 16;
const int rowid = laneid / 16;
......@@ -1761,7 +1744,6 @@ __launch_bounds__(NUM_THREADS, 3) void paged_attention_ll4mi_QKV_mfma16_kernel(
const cache_t* k_ptr2 = k_ptr + kblock_number * kv_block_stride;
const int klocal_token_idx =
TOKENS_PER_WARP * warpid + token_depth * 16 + lane16id;
const int kglobal_token_idx = partition_start_token_idx + klocal_token_idx;
const int kphysical_block_offset = klocal_token_idx % BLOCK_SIZE;
const cache_t* k_ptr3 = k_ptr2 + kphysical_block_offset * KX;
......@@ -2080,9 +2062,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
const int context_len = context_lens[seq_idx];
const int num_partitions = DIVIDE_ROUND_UP(context_len, PARTITION_SIZE);
[[maybe_unused]] constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
const int warpid = threadIdx.x / WARP_SIZE;
[[maybe_unused]] const int laneid = threadIdx.x % WARP_SIZE;
__shared__ float shared_global_exp_sum;
// max num partitions supported is warp_size * NPAR_LOOPS
......@@ -2386,7 +2366,6 @@ __launch_bounds__(NUM_THREADS, 3) void paged_attention_ll4mi_QKV_mfma16_kernel(
const int warpid = threadIdx.x / WARP_SIZE;
const int laneid = threadIdx.x % WARP_SIZE;
const int lane2id = laneid % 2;
const int lane4id = laneid % 4;
const int lane16id = laneid % 16;
const int rowid = laneid / 16;
......@@ -2532,7 +2511,6 @@ __launch_bounds__(NUM_THREADS, 3) void paged_attention_ll4mi_QKV_mfma16_kernel(
const cache_t* k_ptr2 = k_ptr + kblock_number * kv_block_stride;
const int klocal_token_idx =
TOKENS_PER_WARP * warpid + token_depth * 16 + lane16id;
const int kglobal_token_idx = partition_start_token_idx + klocal_token_idx;
const int kphysical_block_offset = klocal_token_idx % BLOCK_SIZE;
const cache_t* k_ptr3 = k_ptr2 + kphysical_block_offset * KX;
......@@ -2816,9 +2794,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
const int context_len = context_lens[seq_idx];
const int num_partitions = DIVIDE_ROUND_UP(context_len, PARTITION_SIZE);
[[maybe_unused]] constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
const int warpid = threadIdx.x / WARP_SIZE;
[[maybe_unused]] const int laneid = threadIdx.x % WARP_SIZE;
__shared__ float shared_global_exp_sum;
// max num partitions supported is warp_size * NPAR_LOOPS
......
csrc/rocm/skinny_gemms.cu
View file @ 99324e25
......@@ -320,7 +320,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
// Goal is to bring the activation matrix A to the LDS
// and use it across the lifetime of the work group
// TODO: When activation matrix is larger than 64 KB
// then this is not goint to work!
// then this is not going to work!
//----------------------------------------------------
__shared__ scalar_t s[max_lds_len];
......@@ -581,7 +581,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
// Goal is to bring the activation matrix A to the LDS
// and use it across the lifetime of the work group
// TODO: When activation matrix is larger than 64 KB
// then this is not goint to work!
// then this is not going to work!
//----------------------------------------------------
__shared__ scalar_t s[max_lds_len];
......@@ -601,7 +601,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
// int _WvPrGrp = mindiv(N, CuCount * YTILE, WvPrGrp);
uint32_t m = (blockIdx.x * _WvPrGrp + threadIdx.y) * YTILE;
// Check whether there will be fragmenation!
// Check whether there will be fragmentation!
// This will happen only for the last wave!
if (m < M && (m + YTILE) >= M) {
uint32_t startColumn = M - YTILE;
......@@ -827,7 +827,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
m += CuCount * _WvPrGrp * YTILE;
// Check whether there will be fragmenation!
// Check whether there will be fragmentation!
// This will happen only for the last wave!
if (m < M && (m + YTILE) >= M) {
uint32_t startColumn = M - YTILE;
......@@ -882,7 +882,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
// Goal is to bring the activation matrix A to the LDS
// and use it across the lifetime of the work group
// TODO: When activation matrix is larger than 64 KB
// then this is not goint to work!
// then this is not going to work!
//----------------------------------------------------
__shared__ scalar_t s[max_lds_len];
......@@ -904,7 +904,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
//----------------------------------------------------
uint32_t m = (blockIdx.x * _WvPrGrp + threadIdx.y) * YTILE;
// Check whether there will be fragmenation!
// Check whether there will be fragmentation!
// This will happen only for the last wave!
if (m < M && (m + YTILE) >= M) {
uint32_t startColumn = M - YTILE;
......@@ -1176,7 +1176,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
m += CuCount * _WvPrGrp * YTILE;
kBase = 0;
// Check whether there will be fragmenation!
// Check whether there will be fragmentation!
// This will happen only for the last wave!
if (m < M && (m + YTILE) >= M) {
uint32_t startColumn = M - YTILE;
......
csrc/sampler.cu
View file @ 99324e25
......@@ -59,6 +59,8 @@ void apply_repetition_penalties_(
int vocab_size = logits.size(-1);
int num_seqs = logits.size(0);
if (num_seqs == 0) return;
// Get number of SMs on the current device
int sms = 0;
cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount,
......
csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu
View file @ 99324e25
......@@ -277,7 +277,7 @@ CompressorResult cutlass_sparse_compress_sm90(torch::Tensor const& a) {
uint32_t const m = 1; // Set M to 1 for compression
uint32_t const n = a.size(1);
// Note: For correctess, the compressed format must be invariant in:
// Note: For correctness, the compressed format must be invariant in:
// - M, the flattened number of tokens
// - Whether output dtype is fp16 or bf16
// - CUTLASS epilogues
......
csrc/sparse/cutlass/sparse_scaled_mm_c3x.cuh
View file @ 99324e25
......@@ -79,7 +79,8 @@ struct cutlass_sparse_3x_gemm {
// These are the minimum alignments needed for the kernels to compile
static constexpr int AlignmentAB =
128 / cutlass::sizeof_bits<ElementAB>::value;
static constexpr int AlignmentCD = 4;
static constexpr int AlignmentCD =
128 / cutlass::sizeof_bits<ElementD>::value;
using CollectiveEpilogue =
typename cutlass::epilogue::collective::CollectiveBuilder<
......
csrc/torch_bindings.cpp
View file @ 99324e25
......@@ -397,6 +397,14 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
{stride_tag});
ops.impl("cutlass_scaled_fp4_mm", torch::kCUDA, &cutlass_scaled_fp4_mm);
// cutlass blockwise scaledgroup GEMM
ops.def(
"cutlass_blockwise_scaled_grouped_mm(Tensor! output, Tensor a, Tensor b, "
"Tensor scales_a, Tensor scales_b, "
"Tensor problem_sizes, Tensor expert_offsets) -> ()",
{stride_tag});
// conditionally compiled so impl registration is in source file
// cutlass nvfp4 block scaled group GEMM
ops.def(
"cutlass_fp4_group_mm(Tensor! out, Tensor a, Tensor b,"
......@@ -729,6 +737,24 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _custom_ar), custom_ar) {
custom_ar.impl("open_mem_handle", torch::kCPU, &open_mem_handle);
custom_ar.def("free_shared_buffer", &free_shared_buffer);
#ifdef USE_ROCM
// Quick Reduce all-reduce kernels
custom_ar.def(
"qr_all_reduce(int fa, Tensor inp, Tensor out, int quant_level, bool "
"cast_bf2half) -> ()");
custom_ar.impl("qr_all_reduce", torch::kCUDA, &qr_all_reduce);
custom_ar.def("init_custom_qr", &init_custom_qr);
custom_ar.def("qr_destroy", &qr_destroy);
custom_ar.def("qr_get_handle", &qr_get_handle);
custom_ar.def("qr_open_handles(int _fa, Tensor[](b!) handles) -> ()");
custom_ar.impl("qr_open_handles", torch::kCPU, &qr_open_handles);
// Max input size in bytes
custom_ar.def("qr_max_size", &qr_max_size);
#endif
}
REGISTER_EXTENSION(TORCH_EXTENSION_NAME)
\ No newline at end of file
docker/Dockerfile
View file @ 99324e25
# The vLLM Dockerfile is used to construct vLLM image that can be directly used
# to run the OpenAI compatible server.
......@@ -6,30 +7,110 @@
# docs/assets/contributing/dockerfile-stages-dependency.png
ARG CUDA_VERSION=12.8.1
ARG PYTHON_VERSION=3.12
# By parameterizing the base images, we allow third-party to use their own
# base images. One use case is hermetic builds with base images stored in
# private registries that use a different repository naming conventions.
#
# Example:
# docker build --build-arg BUILD_BASE_IMAGE=registry.acme.org/mirror/nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04
ARG BUILD_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04
ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
# By parameterizing the Deadsnakes repository URL, we allow third-party to use
# their own mirror. When doing so, we don't benefit from the transparent
# installation of the GPG key of the PPA, as done by add-apt-repository, so we
# also need a URL for the GPG key.
ARG DEADSNAKES_MIRROR_URL
ARG DEADSNAKES_GPGKEY_URL
# The PyPA get-pip.py script is a self contained script+zip file, that provides
# both the installer script and the pip base85-encoded zip archive. This allows
# bootstrapping pip in environment where a dsitribution package does not exist.
#
# By parameterizing the URL for get-pip.py installation script, we allow
# third-party to use their own copy of the script stored in a private mirror.
# We set the default value to the PyPA owned get-pip.py script.
#
# Reference: https://pip.pypa.io/en/stable/installation/#get-pip-py
ARG GET_PIP_URL="https://bootstrap.pypa.io/get-pip.py"
# PIP supports fetching the packages from custom indexes, allowing third-party
# to host the packages in private mirrors. The PIP_INDEX_URL and
# PIP_EXTRA_INDEX_URL are standard PIP environment variables to override the
# default indexes. By letting them empty by default, PIP will use its default
# indexes if the build process doesn't override the indexes.
#
# Uv uses different variables. We set them by default to the same values as
# PIP, but they can be overridden.
ARG PIP_INDEX_URL
ARG PIP_EXTRA_INDEX_URL
ARG UV_INDEX_URL=${PIP_INDEX_URL}
ARG UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
# PyTorch provides its own indexes for standard and nightly builds
ARG PYTORCH_CUDA_INDEX_BASE_URL=https://download.pytorch.org/whl
ARG PYTORCH_CUDA_NIGHTLY_INDEX_BASE_URL=https://download.pytorch.org/whl/nightly
# PIP supports multiple authentication schemes, including keyring
# By parameterizing the PIP_KEYRING_PROVIDER variable and setting it to
# disabled by default, we allow third-party to use keyring authentication for
# their private Python indexes, while not changing the default behavior which
# is no authentication.
#
# Reference: https://pip.pypa.io/en/stable/topics/authentication/#keyring-support
ARG PIP_KEYRING_PROVIDER=disabled
ARG UV_KEYRING_PROVIDER=${PIP_KEYRING_PROVIDER}
# Flag enables build-in KV-connector dependency libs into docker images
ARG INSTALL_KV_CONNECTORS=false
#################### BASE BUILD IMAGE ####################
# prepare basic build environment
FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04 AS base
ARG CUDA_VERSION=12.8.1
ARG PYTHON_VERSION=3.12
FROM ${BUILD_BASE_IMAGE} AS base
ARG CUDA_VERSION
ARG PYTHON_VERSION
ARG TARGETPLATFORM
ARG INSTALL_KV_CONNECTORS=false
ENV DEBIAN_FRONTEND=noninteractive
ARG DEADSNAKES_MIRROR_URL
ARG DEADSNAKES_GPGKEY_URL
ARG GET_PIP_URL
# Install Python and other dependencies
RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
&& echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections \
&& apt-get update -y \
&& apt-get install -y ccache software-properties-common git curl sudo \
&& for i in 1 2 3; do \
add-apt-repository -y ppa:deadsnakes/ppa && break || \
{ echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
done \
&& if [ ! -z ${DEADSNAKES_MIRROR_URL} ] ; then \
if [ ! -z "${DEADSNAKES_GPGKEY_URL}" ] ; then \
mkdir -p -m 0755 /etc/apt/keyrings ; \
curl -L ${DEADSNAKES_GPGKEY_URL} | gpg --dearmor > /etc/apt/keyrings/deadsnakes.gpg ; \
sudo chmod 644 /etc/apt/keyrings/deadsnakes.gpg ; \
echo "deb [signed-by=/etc/apt/keyrings/deadsnakes.gpg] ${DEADSNAKES_MIRROR_URL} $(lsb_release -cs) main" > /etc/apt/sources.list.d/deadsnakes.list ; \
fi ; \
else \
for i in 1 2 3; do \
add-apt-repository -y ppa:deadsnakes/ppa && break || \
{ echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
done ; \
fi \
&& apt-get update -y \
&& apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
&& update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
&& update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
&& ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
&& curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
&& curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION} \
&& python3 --version && python3 -m pip --version
ARG PIP_INDEX_URL UV_INDEX_URL
ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
ARG PYTORCH_CUDA_INDEX_BASE_URL
ARG PYTORCH_CUDA_NIGHTLY_INDEX_BASE_URL
ARG PIP_KEYRING_PROVIDER UV_KEYRING_PROVIDER
# Install uv for faster pip installs
RUN --mount=type=cache,target=/root/.cache/uv \
python3 -m pip install uv
......@@ -63,21 +144,25 @@ WORKDIR /workspace
# after this step
RUN --mount=type=cache,target=/root/.cache/uv \
if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu128 "torch==2.8.0.dev20250318+cu128" "torchvision==0.22.0.dev20250319"; \
uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu128 --pre pytorch_triton==3.3.0+gitab727c40; \
uv pip install --system \
--index-url ${PYTORCH_CUDA_NIGHTLY_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.') \
"torch==2.8.0.dev20250318+cu128" "torchvision==0.22.0.dev20250319"; \
uv pip install --system \
--index-url ${PYTORCH_CUDA_NIGHTLY_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.') \
--pre pytorch_triton==3.3.0+gitab727c40; \
fi
COPY requirements/common.txt requirements/common.txt
COPY requirements/cuda.txt requirements/cuda.txt
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system -r requirements/cuda.txt \
--extra-index-url https://download.pytorch.org/whl/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
# cuda arch list used by torch
# can be useful for both `dev` and `test`
# explicitly set the list to avoid issues with torch 2.2
# see https://github.com/pytorch/pytorch/pull/123243
ARG torch_cuda_arch_list='7.0 7.5 8.0 8.9 9.0 10.0+PTX'
ARG torch_cuda_arch_list='7.0 7.5 8.0 8.9 9.0 10.0 12.0'
ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
# Override the arch list for flash-attn to reduce the binary size
ARG vllm_fa_cmake_gpu_arches='80-real;90-real'
......@@ -88,6 +173,10 @@ ENV VLLM_FA_CMAKE_GPU_ARCHES=${vllm_fa_cmake_gpu_arches}
FROM base AS build
ARG TARGETPLATFORM
ARG PIP_INDEX_URL UV_INDEX_URL
ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
ARG PYTORCH_CUDA_INDEX_BASE_URL
# install build dependencies
COPY requirements/build.txt requirements/build.txt
......@@ -98,7 +187,7 @@ ENV UV_INDEX_STRATEGY="unsafe-best-match"
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system -r requirements/build.txt \
--extra-index-url https://download.pytorch.org/whl/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
COPY . .
ARG GIT_REPO_CHECK=0
......@@ -113,6 +202,8 @@ ARG nvcc_threads=8
ENV NVCC_THREADS=$nvcc_threads
ARG USE_SCCACHE
ARG SCCACHE_DOWNLOAD_URL=https://github.com/mozilla/sccache/releases/download/v0.8.1/sccache-v0.8.1-x86_64-unknown-linux-musl.tar.gz
ARG SCCACHE_ENDPOINT
ARG SCCACHE_BUCKET_NAME=vllm-build-sccache
ARG SCCACHE_REGION_NAME=us-west-2
ARG SCCACHE_S3_NO_CREDENTIALS=0
......@@ -121,10 +212,11 @@ RUN --mount=type=cache,target=/root/.cache/uv \
--mount=type=bind,source=.git,target=.git \
if [ "$USE_SCCACHE" = "1" ]; then \
echo "Installing sccache..." \
&& curl -L -o sccache.tar.gz https://github.com/mozilla/sccache/releases/download/v0.8.1/sccache-v0.8.1-x86_64-unknown-linux-musl.tar.gz \
&& curl -L -o sccache.tar.gz ${SCCACHE_DOWNLOAD_URL} \
&& tar -xzf sccache.tar.gz \
&& sudo mv sccache-v0.8.1-x86_64-unknown-linux-musl/sccache /usr/bin/sccache \
&& rm -rf sccache.tar.gz sccache-v0.8.1-x86_64-unknown-linux-musl \
&& if [ ! -z ${SCCACHE_ENDPOINT} ] ; then export SCCACHE_ENDPOINT=${SCCACHE_ENDPOINT} ; fi \
&& export SCCACHE_BUCKET=${SCCACHE_BUCKET_NAME} \
&& export SCCACHE_REGION=${SCCACHE_REGION_NAME} \
&& export SCCACHE_S3_NO_CREDENTIALS=${SCCACHE_S3_NO_CREDENTIALS} \
......@@ -162,6 +254,10 @@ RUN if [ "$RUN_WHEEL_CHECK" = "true" ]; then \
#################### DEV IMAGE ####################
FROM base as dev
ARG PIP_INDEX_URL UV_INDEX_URL
ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
ARG PYTORCH_CUDA_INDEX_BASE_URL
# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
# Reference: https://github.com/astral-sh/uv/pull/1694
ENV UV_HTTP_TIMEOUT=500
......@@ -176,21 +272,26 @@ COPY requirements/test.txt requirements/test.txt
COPY requirements/dev.txt requirements/dev.txt
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system -r requirements/dev.txt \
--extra-index-url https://download.pytorch.org/whl/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
#################### DEV IMAGE ####################
#################### vLLM installation IMAGE ####################
# image with vLLM installed
# TODO: Restore to base image after FlashInfer AOT wheel fixed
FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS vllm-base
ARG CUDA_VERSION=12.8.1
ARG PYTHON_VERSION=3.12
FROM ${FINAL_BASE_IMAGE} AS vllm-base
ARG CUDA_VERSION
ARG PYTHON_VERSION
ARG INSTALL_KV_CONNECTORS=false
WORKDIR /vllm-workspace
ENV DEBIAN_FRONTEND=noninteractive
ARG TARGETPLATFORM
SHELL ["/bin/bash", "-c"]
ARG DEADSNAKES_MIRROR_URL
ARG DEADSNAKES_GPGKEY_URL
ARG GET_PIP_URL
RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
......@@ -200,17 +301,33 @@ RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
&& apt-get update -y \
&& apt-get install -y ccache software-properties-common git curl wget sudo vim python3-pip \
&& apt-get install -y ffmpeg libsm6 libxext6 libgl1 \
&& for i in 1 2 3; do \
add-apt-repository -y ppa:deadsnakes/ppa && break || \
{ echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
done \
&& if [ ! -z ${DEADSNAKES_MIRROR_URL} ] ; then \
if [ ! -z "${DEADSNAKES_GPGKEY_URL}" ] ; then \
mkdir -p -m 0755 /etc/apt/keyrings ; \
curl -L ${DEADSNAKES_GPGKEY_URL} | gpg --dearmor > /etc/apt/keyrings/deadsnakes.gpg ; \
sudo chmod 644 /etc/apt/keyrings/deadsnakes.gpg ; \
echo "deb [signed-by=/etc/apt/keyrings/deadsnakes.gpg] ${DEADSNAKES_MIRROR_URL} $(lsb_release -cs) main" > /etc/apt/sources.list.d/deadsnakes.list ; \
fi ; \
else \
for i in 1 2 3; do \
add-apt-repository -y ppa:deadsnakes/ppa && break || \
{ echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
done ; \
fi \
&& apt-get update -y \
&& apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv libibverbs-dev \
&& update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
&& update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
&& ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
&& curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
&& curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION} \
&& python3 --version && python3 -m pip --version
ARG PIP_INDEX_URL UV_INDEX_URL
ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
ARG PYTORCH_CUDA_INDEX_BASE_URL
ARG PYTORCH_CUDA_NIGHTLY_INDEX_BASE_URL
ARG PIP_KEYRING_PROVIDER UV_KEYRING_PROVIDER
# Install uv for faster pip installs
RUN --mount=type=cache,target=/root/.cache/uv \
python3 -m pip install uv
......@@ -232,43 +349,75 @@ RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
# after this step
RUN --mount=type=cache,target=/root/.cache/uv \
if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu128 "torch==2.8.0.dev20250318+cu128" "torchvision==0.22.0.dev20250319"; \
uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu128 --pre pytorch_triton==3.3.0+gitab727c40; \
uv pip install --system \
--index-url ${PYTORCH_CUDA_NIGHTLY_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.') \
"torch==2.8.0.dev20250318+cu128" "torchvision==0.22.0.dev20250319" ; \
uv pip install --system \
--index-url ${PYTORCH_CUDA_NIGHTLY_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.') \
--pre pytorch_triton==3.3.0+gitab727c40 ; \
fi
# Install vllm wheel first, so that torch etc will be installed.
RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist \
--mount=type=cache,target=/root/.cache/uv \
uv pip install --system dist/*.whl --verbose \
--extra-index-url https://download.pytorch.org/whl/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
# If we need to build FlashInfer wheel before its release:
# $ export FLASHINFER_ENABLE_AOT=1
# $ # Note we remove 7.0 from the arch list compared to the list below, since FlashInfer only supports sm75+
# $ export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.6 8.9 9.0+PTX'
# $ export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.9 9.0a 10.0a 12.0'
# $ git clone https://github.com/flashinfer-ai/flashinfer.git --recursive
# $ cd flashinfer
# $ git checkout 524304395bd1d8cd7d07db083859523fcaa246a4
# $ rm -rf build
# $ python3 setup.py bdist_wheel --dist-dir=dist --verbose
# $ ls dist
# $ # upload the wheel to a public location, e.g. https://wheels.vllm.ai/flashinfer/524304395bd1d8cd7d07db083859523fcaa246a4/flashinfer_python-0.2.1.post1+cu124torch2.5-cp38-abi3-linux_x86_64.whl
RUN --mount=type=cache,target=/root/.cache/uv \
. /etc/environment && \
if [ "$TARGETPLATFORM" != "linux/arm64" ]; then \
# FlashInfer alreary has a wheel for PyTorch 2.7.0 and CUDA 12.8. This is enough for CI use
if [[ "$CUDA_VERSION" == 12.8* ]]; then \
uv pip install --system https://download.pytorch.org/whl/cu128/flashinfer/flashinfer_python-0.2.5%2Bcu128torch2.7-cp38-abi3-linux_x86_64.whl; \
else \
export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.9 9.0+PTX'; \
CUDA_MAJOR="${CUDA_VERSION%%.*}"; \
if [ "$CUDA_MAJOR" -lt 12 ]; then \
export FLASHINFER_ENABLE_SM90=0; \
fi; \
uv pip install --system --no-build-isolation "git+https://github.com/flashinfer-ai/flashinfer@21ea1d2545f74782b91eb8c08fd503ac4c0743fc" ; \
fi \
fi
# $ git checkout v0.2.6.post1
# $ python -m flashinfer.aot
# $ python -m build --no-isolation --wheel
# $ ls -la dist
# -rw-rw-r-- 1 mgoin mgoin 205M Jun 9 18:03 flashinfer_python-0.2.6.post1-cp39-abi3-linux_x86_64.whl
# $ # upload the wheel to a public location, e.g. https://wheels.vllm.ai/flashinfer/v0.2.6.post1/flashinfer_python-0.2.6.post1-cp39-abi3-linux_x86_64.whl
# Allow specifying a version, Git revision or local .whl file
ARG FLASHINFER_CUDA128_INDEX_URL="https://download.pytorch.org/whl/cu128/flashinfer"
ARG FLASHINFER_CUDA128_WHEEL="flashinfer_python-0.2.6.post1%2Bcu128torch2.7-cp39-abi3-linux_x86_64.whl"
ARG FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
ARG FLASHINFER_GIT_REF="v0.2.6.post1"
RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
. /etc/environment
if [ "$TARGETPLATFORM" != "linux/arm64" ]; then
# FlashInfer already has a wheel for PyTorch 2.7.0 and CUDA 12.8. This is enough for CI use
if [[ "$CUDA_VERSION" == 12.8* ]]; then
uv pip install --system ${FLASHINFER_CUDA128_INDEX_URL}/${FLASHINFER_CUDA128_WHEEL}
else
export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.9 9.0a 10.0a 12.0'
git clone ${FLASHINFER_GIT_REPO} --single-branch --branch ${FLASHINFER_GIT_REF} --recursive
# Needed to build AOT kernels
(cd flashinfer && \
python3 -m flashinfer.aot && \
uv pip install --system --no-build-isolation . \
)
rm -rf flashinfer
# Default arches (skipping 10.0a and 12.0 since these need 12.8)
# TODO: Update this to allow setting TORCH_CUDA_ARCH_LIST as a build arg.
TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9 9.0a"
if [[ "${CUDA_VERSION}" == 11.* ]]; then
TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9"
fi
echo "🏗️ Building FlashInfer for arches: ${TORCH_CUDA_ARCH_LIST}"
git clone --depth 1 --recursive --shallow-submodules \
--branch v0.2.6.post1 \
https://github.com/flashinfer-ai/flashinfer.git flashinfer
pushd flashinfer
python3 -m flashinfer.aot
TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST}" \
uv pip install --system --no-build-isolation .
popd
rm -rf flashinfer
fi \
fi
BASH
COPY examples examples
COPY benchmarks benchmarks
COPY ./vllm/collect_env.py .
......@@ -284,7 +433,7 @@ uv pip list
COPY requirements/build.txt requirements/build.txt
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system -r requirements/build.txt \
--extra-index-url https://download.pytorch.org/whl/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
#################### vLLM installation IMAGE ####################
......@@ -295,6 +444,11 @@ FROM vllm-base AS test
ADD . /vllm-workspace/
ARG PYTHON_VERSION
ARG PIP_INDEX_URL UV_INDEX_URL
ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
# Reference: https://github.com/astral-sh/uv/pull/1694
ENV UV_HTTP_TIMEOUT=500
......@@ -305,7 +459,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system --no-build-isolation "git+https://github.com/state-spaces/mamba@v2.2.4"
# install development dependencies (for testing)
RUN --mount=type=cache,target=/root/.cache/uv \
RUN --mount=type=cache,target=/root/.cache/uv \
CUDA_MAJOR="${CUDA_VERSION%%.*}"; \
if [ "$CUDA_MAJOR" -ge 12 ]; then \
uv pip install --system -r requirements/dev.txt; \
......@@ -321,7 +475,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
ENV HF_HUB_ENABLE_HF_TRANSFER 1
# Copy in the v1 package for testing (it isn't distributed yet)
COPY vllm/v1 /usr/local/lib/python3.12/dist-packages/vllm/v1
COPY vllm/v1 /usr/local/lib/python${PYTHON_VERSION}/dist-packages/vllm/v1
# doc requires source code
# we hide them inside `test_docs/` , so that this source code
......@@ -337,17 +491,26 @@ RUN mv mkdocs.yaml test_docs/
# base openai image with additional requirements, for any subsequent openai-style images
FROM vllm-base AS vllm-openai-base
ARG TARGETPLATFORM
ARG INSTALL_KV_CONNECTORS=false
ARG PIP_INDEX_URL UV_INDEX_URL
ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
# Reference: https://github.com/astral-sh/uv/pull/1694
ENV UV_HTTP_TIMEOUT=500
COPY requirements/kv_connectors.txt requirements/kv_connectors.txt
# install additional dependencies for openai api server
RUN --mount=type=cache,target=/root/.cache/uv \
if [ "$INSTALL_KV_CONNECTORS" = "true" ]; then \
uv pip install --system -r requirements/kv_connectors.txt; \
fi; \
if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
uv pip install --system accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.42.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
else \
uv pip install --system accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.45.3' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
uv pip install --system accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.46.1' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
fi
ENV VLLM_USAGE_SOURCE production-docker-image
......
docker/Dockerfile.cpu
View file @ 99324e25
......@@ -8,6 +8,8 @@
# Build arguments:
# PYTHON_VERSION=3.12 (default)|3.11|3.10|3.9
# VLLM_CPU_DISABLE_AVX512=false (default)|true
# VLLM_CPU_AVX512BF16=false (default)|true
# VLLM_CPU_AVX512VNNI=false (default)|true
#
######################### BASE IMAGE #########################
......@@ -25,7 +27,7 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
--mount=type=cache,target=/var/lib/apt,sharing=locked \
apt-get update -y \
&& apt-get install -y --no-install-recommends ccache git curl wget ca-certificates \
gcc-12 g++-12 libtcmalloc-minimal4 libnuma-dev ffmpeg libsm6 libxext6 libgl1 \
gcc-12 g++-12 libtcmalloc-minimal4 libnuma-dev ffmpeg libsm6 libxext6 libgl1 jq lsof \
&& update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12 \
&& curl -LsSf https://astral.sh/uv/install.sh | sh
......@@ -60,13 +62,19 @@ FROM base AS vllm-build
ARG GIT_REPO_CHECK=0
# Support for building with non-AVX512 vLLM: docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" ...
ARG VLLM_CPU_DISABLE_AVX512
ARG VLLM_CPU_DISABLE_AVX512=0
ENV VLLM_CPU_DISABLE_AVX512=${VLLM_CPU_DISABLE_AVX512}
# Support for building with AVX512BF16 ISA: docker build --build-arg VLLM_CPU_AVX512BF16="true" ...
ARG VLLM_CPU_AVX512BF16=0
ENV VLLM_CPU_AVX512BF16=${VLLM_CPU_AVX512BF16}
# Support for building with AVX512VNNI ISA: docker build --build-arg VLLM_CPU_AVX512VNNI="true" ...
ARG VLLM_CPU_AVX512VNNI=0
ENV VLLM_CPU_AVX512VNNI=${VLLM_CPU_AVX512VNNI}
WORKDIR /workspace/vllm
RUN --mount=type=cache,target=/root/.cache/uv \
--mount=type=bind,src=requirements/build.txt,target=requirements/build.txt \
--mount=type=bind,src=requirements/cpu-build.txt,target=requirements/build.txt \
uv pip install -r requirements/build.txt
COPY . .
......@@ -79,6 +87,22 @@ RUN --mount=type=cache,target=/root/.cache/uv \
--mount=type=bind,source=.git,target=.git \
VLLM_TARGET_DEVICE=cpu python3 setup.py bdist_wheel
######################### TEST DEPS #########################
FROM base AS vllm-test-deps
WORKDIR /workspace/vllm
RUN --mount=type=bind,src=requirements/test.in,target=requirements/test.in \
cp requirements/test.in requirements/cpu-test.in && \
sed -i '/mamba_ssm/d' requirements/cpu-test.in && \
sed -i 's/torch==.*/torch==2.6.0/g' requirements/cpu-test.in && \
sed -i 's/torchaudio.*/torchaudio/g' requirements/cpu-test.in && \
sed -i 's/torchvision.*/torchvision/g' requirements/cpu-test.in && \
uv pip compile requirements/cpu-test.in -o requirements/cpu-test.txt --index-strategy unsafe-best-match --torch-backend cpu
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install -r requirements/cpu-test.txt
######################### DEV IMAGE #########################
FROM vllm-build AS vllm-dev
......@@ -97,6 +121,8 @@ RUN --mount=type=cache,target=/root/.cache/uv \
--mount=type=bind,source=.git,target=.git \
VLLM_TARGET_DEVICE=cpu python3 setup.py develop
COPY --from=vllm-test-deps /workspace/vllm/requirements/cpu-test.txt requirements/test.txt
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install -r requirements/dev.txt && \
pre-commit install --hook-type pre-commit --hook-type commit-msg
......@@ -104,17 +130,10 @@ RUN --mount=type=cache,target=/root/.cache/uv \
ENTRYPOINT ["bash"]
######################### TEST IMAGE #########################
FROM base AS vllm-test
FROM vllm-test-deps AS vllm-test
WORKDIR /workspace/
RUN --mount=type=cache,target=/root/.cache/uv \
--mount=type=bind,src=requirements/test.in,target=requirements/test.in \
cp requirements/test.in requirements/test-cpu.in && \
sed -i '/mamba_ssm/d' requirements/test-cpu.in && \
uv pip compile requirements/test-cpu.in -o requirements/cpu-test.txt && \
uv pip install -r requirements/cpu-test.txt
RUN --mount=type=cache,target=/root/.cache/uv \
--mount=type=bind,from=vllm-build,src=/workspace/vllm/dist,target=dist \
uv pip install dist/*.whl
......@@ -123,6 +142,7 @@ ADD ./tests/ ./tests/
ADD ./examples/ ./examples/
ADD ./benchmarks/ ./benchmarks/
ADD ./vllm/collect_env.py .
ADD ./.buildkite/ ./.buildkite/
# install development dependencies (for testing)
RUN --mount=type=cache,target=/root/.cache/uv \
......
docker/Dockerfile.rocm_base
View file @ 99324e25
......@@ -12,7 +12,7 @@ ARG PYTORCH_REPO="https://github.com/pytorch/pytorch.git"
ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git"
ARG FA_BRANCH="1a7f4dfa"
ARG FA_REPO="https://github.com/Dao-AILab/flash-attention.git"
ARG AITER_BRANCH="c1debd8"
ARG AITER_BRANCH="6487649"
ARG AITER_REPO="https://github.com/ROCm/aiter.git"
FROM ${BASE_IMAGE} AS base
......
docker/Dockerfile.xpu
View file @ 99324e25
......@@ -35,6 +35,7 @@ RUN --mount=type=bind,source=.git,target=.git \
if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh; fi
ENV VLLM_TARGET_DEVICE=xpu
ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
RUN --mount=type=cache,target=/root/.cache/pip \
--mount=type=bind,source=.git,target=.git \
......
docs/.nav.yml
View file @ 99324e25
......@@ -39,6 +39,7 @@ nav:
- models/generative_models.md
- models/pooling_models.md
- models/extensions
- Hardware Supported Models: models/hardware_supported_models
- Features:
- features/compatibility_matrix.md
- features/*
......@@ -48,7 +49,12 @@ nav:
- General:
- glob: contributing/*
flatten_single_child_sections: true
- Model Implementation: contributing/model
- Model Implementation:
- contributing/model/README.md
- contributing/model/basic.md
- contributing/model/registration.md
- contributing/model/tests.md
- contributing/model/multimodal.md
- Design Documents:
- V0: design
- V1: design/v1
......
docs/README.md
View file @ 99324e25
# Welcome to vLLM
<figure markdown="span">
![](./assets/logos/vllm-logo-text-light.png){ align="center" alt="vLLM" class="no-scaled-link" width="60%" }
![](./assets/logos/vllm-logo-text-light.png){ align="center" alt="vLLM Light" class="logo-light" width="60%" }
![](./assets/logos/vllm-logo-text-dark.png){ align="center" alt="vLLM Dark" class="logo-dark" width="60%" }
</figure>
<p style="text-align:center">
......@@ -40,7 +41,7 @@ vLLM is flexible and easy to use with:
- OpenAI-compatible API server
- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs, Gaudi® accelerators and GPUs, IBM Power CPUs, TPU, and AWS Trainium and Inferentia Accelerators.
- Prefix caching support
- Multi-lora support
- Multi-LoRA support
For more information, check out the following:
......
docs/ci/update_pytorch_version.md 0 → 100644
View file @ 99324e25
---
title: Update PyTorch version on vLLM OSS CI/CD
---
vLLM's current policy is to always use the latest PyTorch stable
release in CI/CD. It is standard practice to submit a PR to update the
PyTorch version as early as possible when a new [PyTorch stable
release](https://github.com/pytorch/pytorch/blob/main/RELEASE.md#release-cadence) becomes available.
This process is non-trivial due to the gap between PyTorch
releases. Using [#16859](https://github.com/vllm-project/vllm/pull/16859) as
an example, this document outlines common steps to achieve this update along with
a list of potential issues and how to address them.
## Test PyTorch release candidates (RCs)
Updating PyTorch in vLLM after the official release is not
ideal because any issues discovered at that point can only be resolved
by waiting for the next release or by implementing hacky workarounds in vLLM.
The better solution is to test vLLM with PyTorch release candidates (RC) to ensure
compatibility before each release.
PyTorch release candidates can be downloaded from PyTorch test index at https://download.pytorch.org/whl/test.
For example, torch2.7.0+cu12.8 RC can be installed using the following command:
```
uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/test/cu128
```
When the final RC is ready for testing, it will be announced to the community
on the [PyTorch dev-discuss forum](https://dev-discuss.pytorch.org/c/release-announcements).
After this announcement, we can begin testing vLLM integration by drafting a pull request
following this 3-step process:
1. Update requirements files in https://github.com/vllm-project/vllm/tree/main/requirements
to point to the new releases for torch, torchvision, and torchaudio.
2. Use `--extra-index-url https://download.pytorch.org/whl/test/<PLATFORM>` to
get the final release candidates' wheels. Some common platforms are `cpu`, `cu128`,
and `rocm6.2.4`.
3. As vLLM uses uv, make sure that `unsafe-best-match` strategy is set either
via `UV_INDEX_STRATEGY` env variable or via `--index-strategy unsafe-best-match`.
If failures are found in the pull request, raise them as issues on vLLM and
cc the PyTorch release team to initiate discussion on how to address them.
## Update CUDA version
The PyTorch release matrix includes both stable and experimental [CUDA versions](https://github.com/pytorch/pytorch/blob/main/RELEASE.md#release-compatibility-matrix). Due to limitations, only the latest stable CUDA version (for example,
torch2.7.0+cu12.6) is uploaded to PyPI. However, vLLM may require a different CUDA version,
such as 12.8 for Blackwell support.
This complicates the process as we cannot use the out-of-the-box
`pip install torch torchvision torchaudio` command. The solution is to use
`--extra-index-url` in vLLM's Dockerfiles.
1. Use `--extra-index-url https://download.pytorch.org/whl/cu128` to install torch+cu128.
2. Other important indexes at the moment include:
1. CPU ‒ https://download.pytorch.org/whl/cpu
2. ROCm ‒ https://download.pytorch.org/whl/rocm6.2.4 and https://download.pytorch.org/whl/rocm6.3
3. XPU ‒ https://download.pytorch.org/whl/xpu
3. Update .buildkite/release-pipeline.yaml and .buildkite/scripts/upload-wheels.sh to
match the CUDA version from step 1. This makes sure that the release vLLM wheel is tested
on CI.
## Address long vLLM build time
When building vLLM with a new PyTorch/CUDA version, no cache will exist
in the vLLM sccache S3 bucket, causing the build job on CI to potentially take more than 5 hours
and timeout. Additionally, since vLLM's fastcheck pipeline runs in read-only mode,
it doesn't populate the cache, so re-running it to warm up the cache
is ineffective.
While ongoing efforts like [#17419](https://github.com/vllm-project/vllm/issues/17419)
address the long build time at its source, the current workaround is to set VLLM_CI_BRANCH
to a custom branch provided by @khluu (`VLLM_CI_BRANCH=khluu/use_postmerge_q`)
when manually triggering a build on Buildkite. This branch accomplishes two things:
1. Increase the timeout limit to 10 hours so that the build doesn't timeout.
2. Allow the compiled artifacts to be written to the vLLM sccache S3 bucket
to warm it up so that future builds are faster.
<p align="center" width="100%">
<img width="60%" src="https://github.com/user-attachments/assets/a8ff0fcd-76e0-4e91-b72f-014e3fdb6b94">
</p>
## Update dependencies
Several vLLM dependencies, such as FlashInfer, also depend on PyTorch and need
to be updated accordingly. Rather than waiting for all of them to publish new
releases (which would take too much time), they can be built from
source to unblock the update process.
### FlashInfer
Here is how to build and install it from source with torch2.7.0+cu128 in vLLM [Dockerfile](https://github.com/vllm-project/vllm/blob/27bebcd89792d5c4b08af7a65095759526f2f9e1/docker/Dockerfile#L259-L271):
```bash
export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.9 9.0 10.0+PTX'
export FLASHINFER_ENABLE_SM90=1
uv pip install --system --no-build-isolation "git+https://github.com/flashinfer-ai/flashinfer@v0.2.6.post1"
```
One caveat is that building FlashInfer from source adds approximately 30
minutes to the vLLM build time. Therefore, it's preferable to cache the wheel in a
public location for immediate installation, such as https://download.pytorch.org/whl/cu128/flashinfer/flashinfer_python-0.2.6.post1%2Bcu128torch2.7-cp39-abi3-linux_x86_64.whl. For future releases, contact the PyTorch release
team if you want to get the package published there.
### xFormers
Similar to FlashInfer, here is how to build and install xFormers from source:
```bash
export TORCH_CUDA_ARCH_LIST='7.0 7.5 8.0 8.9 9.0 10.0+PTX'
MAX_JOBS=16 uv pip install --system --no-build-isolation "git+https://github.com/facebookresearch/xformers@v0.0.30"
```
### Mamba
```bash
uv pip install --system --no-build-isolation "git+https://github.com/state-spaces/mamba@v2.2.4"
```
### causal-conv1d
```
uv pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
```
## Update all the different vLLM platforms
Rather than attempting to update all vLLM platforms in a single pull request, it's more manageable
to handle some platforms separately. The separation of requirements and Dockerfiles
for different platforms in vLLM CI/CD allows us to selectively choose
which platforms to update. For instance, updating XPU requires the corresponding
release from https://github.com/intel/intel-extension-for-pytorch by Intel.
While https://github.com/vllm-project/vllm/pull/16859 updated vLLM to PyTorch
2.7.0 on CPU, CUDA, and ROCm, https://github.com/vllm-project/vllm/pull/17444
completed the update for XPU.
docs/cli/README.md
View file @ 99324e25
......@@ -16,35 +16,33 @@ vllm {chat,complete,serve,bench,collect-env,run-batch}
Start the vLLM OpenAI Compatible API server.
Examples:
??? Examples
```bash
# Start with a model
vllm serve meta-llama/Llama-2-7b-hf
```bash
# Start with a model
vllm serve meta-llama/Llama-2-7b-hf
# Specify the port
vllm serve meta-llama/Llama-2-7b-hf --port 8100
# Specify the port
vllm serve meta-llama/Llama-2-7b-hf --port 8100
# Check with --help for more options
# To list all groups
vllm serve --help=listgroup
# Check with --help for more options
# To list all groups
vllm serve --help=listgroup
# To view a argument group
vllm serve --help=ModelConfig
# To view a argument group
vllm serve --help=ModelConfig
# To view a single argument
vllm serve --help=max-num-seqs
# To view a single argument
vllm serve --help=max-num-seqs
# To search by keyword
vllm serve --help=max
```
# To search by keyword
vllm serve --help=max
```
## chat
Generate chat completions via the running API server.
Examples:
```bash
# Directly connect to localhost API without arguments
vllm chat
......@@ -60,8 +58,6 @@ vllm chat --quick "hi"
Generate text completions based on the given prompt via the running API server.
Examples:
```bash
# Directly connect to localhost API without arguments
vllm complete
......@@ -73,6 +69,8 @@ vllm complete --url http://{vllm-serve-host}:{vllm-serve-port}/v1
vllm complete --quick "The future of AI is"
```
</details>
## bench
Run benchmark tests for latency online serving throughput and offline inference throughput.
......@@ -89,8 +87,6 @@ vllm bench {latency, serve, throughput}
Benchmark the latency of a single batch of requests.
Example:
```bash
vllm bench latency \
--model meta-llama/Llama-3.2-1B-Instruct \
......@@ -104,8 +100,6 @@ vllm bench latency \
Benchmark the online serving throughput.
Example:
```bash
vllm bench serve \
--model meta-llama/Llama-3.2-1B-Instruct \
......@@ -120,8 +114,6 @@ vllm bench serve \
Benchmark offline inference throughput.
Example:
```bash
vllm bench throughput \
--model meta-llama/Llama-3.2-1B-Instruct \
......@@ -143,7 +135,8 @@ vllm collect-env
Run batch prompts and write results to file.
Examples:
<details>
<summary>Examples</summary>
```bash
# Running with a local file
......@@ -159,6 +152,8 @@ vllm run-batch \
--model meta-llama/Meta-Llama-3-8B-Instruct
```
</details>
## More Help
For detailed options of any subcommand, use:
......
docs/community/contact_us.md 0 → 100644
View file @ 99324e25
---
title: Contact Us
---
[](){ #contactus }
--8<-- "README.md:contact-us"
docs/configuration/conserving_memory.md
View file @ 99324e25
......@@ -57,19 +57,21 @@ By default, we optimize model inference using CUDA graphs which take up extra me
You can adjust `compilation_config` to achieve a better balance between inference speed and memory usage:
```python
from vllm import LLM
from vllm.config import CompilationConfig, CompilationLevel
llm = LLM(
model="meta-llama/Llama-3.1-8B-Instruct",
compilation_config=CompilationConfig(
level=CompilationLevel.PIECEWISE,
# By default, it goes up to max_num_seqs
cudagraph_capture_sizes=[1, 2, 4, 8, 16],
),
)
```
??? Code
```python
from vllm import LLM
from vllm.config import CompilationConfig, CompilationLevel
llm = LLM(
model="meta-llama/Llama-3.1-8B-Instruct",
compilation_config=CompilationConfig(
level=CompilationLevel.PIECEWISE,
# By default, it goes up to max_num_seqs
cudagraph_capture_sizes=[1, 2, 4, 8, 16],
),
)
```
You can disable graph capturing completely via the `enforce_eager` flag:
......@@ -127,18 +129,20 @@ reduce the size of the processed multi-modal inputs, which in turn saves memory.
Here are some examples:
```python
from vllm import LLM
??? Code
# Available for Qwen2-VL series models
llm = LLM(model="Qwen/Qwen2.5-VL-3B-Instruct",
mm_processor_kwargs={
"max_pixels": 768 * 768, # Default is 1280 * 28 * 28
})
# Available for InternVL series models
llm = LLM(model="OpenGVLab/InternVL2-2B",
mm_processor_kwargs={
"max_dynamic_patch": 4, # Default is 12
})
```
```python
from vllm import LLM
# Available for Qwen2-VL series models
llm = LLM(model="Qwen/Qwen2.5-VL-3B-Instruct",
mm_processor_kwargs={
"max_pixels": 768 * 768, # Default is 1280 * 28 * 28
})
# Available for InternVL series models
llm = LLM(model="OpenGVLab/InternVL2-2B",
mm_processor_kwargs={
"max_dynamic_patch": 4, # Default is 12
})
```
docs/configuration/engine_args.md
View file @ 99324e25
......@@ -6,7 +6,7 @@ title: Engine Arguments
Engine arguments control the behavior of the vLLM engine.
- For [offline inference][offline-inference], they are part of the arguments to [LLM][vllm.LLM] class.
- For [online serving][openai-compatible-server], they are part of the arguments to `vllm serve`.
- For [online serving][serving-openai-compatible-server], they are part of the arguments to `vllm serve`.
You can look at [EngineArgs][vllm.engine.arg_utils.EngineArgs] and [AsyncEngineArgs][vllm.engine.arg_utils.AsyncEngineArgs] to see the available engine arguments.
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment