Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

llama : add high-throughput mode (#14363)

Browse Source 
* kv-cache : prepare K/V buffers for separation

ggml-ci

* batched-bench : fix oob write

ggml-ci

* llama : add "virtual sequences"

ggml-ci

* llama : use "stream" vs "virtual sequence"

ggml-ci

* graph : fix stream splitting when KV cache is not used

ggml-ci

* kv-cache : add multi-stream save/load support

ggml-ci

* llama : add "--attn-streams" flag

ggml-ci

* kv-cache : fix handling when find_slot fails

ggml-ci

* kv-cache : restore find_slot impl

ggml-ci

* kv-cache : add comments

* kv-cache : add bounds checks for sequence id

ggml-ci

* cont : add n_seq_max to batch allocr

ggml-ci

* kv-cache : perform stream copies lazily after llama_synchronize

ggml-ci

* kv-cache : avoid throwing exceptions across the C boundary

ggml-ci

* CUDA: 4D FlashAttention support (#14628)

* CUDA: 4D FlashAttention support

* CUDA: fix WMMA FA kernel

* llama : rename attn_streams -> kv_unified

ggml-ci

* common : rename kv_split -> kv_unified

ggml-ci

---------

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
This commit is contained in:
Georgi Gerganov
2025-07-16 16:35:42 +03:00
committed by GitHub
co-authored by Johannes Gäßler
parent ab14019821
commit 225e7a1438
30 changed files with 1080 additions and 460 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/fattn-tile-f32.cu
+19 -13
View File
@@ -31,8 +31,10 @@ static __global__ void flash_attn_tile_ext_f32(
const int ne13,
const int ne31,
const int ne32,
const int ne33,
const int nb31,
const int nb32,
const int nb33,
const int nb01,
const int nb02,
const int nb03,
@@ -74,15 +76,17 @@ static __global__ void flash_attn_tile_ext_f32(
const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
const int sequence = blockIdx.z / ne02;
const int head = blockIdx.z - sequence*ne02;
const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
const float2 * Q_f2 = (const float2 *) (Q + nb02* blockIdx.z + nb01*ic0);
const half2 * K_h2 = (const half2 *) (K + nb12*(blockIdx.z / gqa_ratio));
const half2 * V_h2 = (const half2 *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
const float2 * Q_f2 = (const float2 *) (Q + nb03* sequence + nb02* head + nb01*ic0);
const half2 * K_h2 = (const half2 *) (K + nb13* sequence + nb12*(head / gqa_ratio));
const half2 * V_h2 = (const half2 *) (V + nb13* sequence + nb12*(head / gqa_ratio)); // K and V have same shape
const half * maskh = (const half *) (mask + nb33*(sequence % ne33) + nb31*ic0);
const int stride_KV2 = nb11 / sizeof(half2);
const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
const float slope = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
@@ -265,6 +269,8 @@ static __global__ void flash_attn_tile_ext_f32(
__syncthreads();
}
float2 * dst2 = (float2 *) dst;
#pragma unroll
for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
const int j_VKQ = j_VKQ_0 + threadIdx.y;
@@ -276,22 +282,22 @@ static __global__ void flash_attn_tile_ext_f32(
float kqsum_j = kqsum[j_VKQ_0/nwarps];
kqsum_j = warp_reduce_sum(kqsum_j);
#pragma unroll
for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
const int i0 = i00 + 2*threadIdx.x;
const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
#pragma unroll
for (int i00 = 0; i00 < D/2; i00 += WARP_SIZE) {
const int i0 = i00 + threadIdx.x;
float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/WARP_SIZE];
if (gridDim.y == 1) {
dst_val.x /= kqsum_j;
dst_val.y /= kqsum_j;
}
const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 0] = dst_val.x;
dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 1] = dst_val.y;
dst2[j_dst_unrolled*(D/2) + i0] = dst_val;
}
if (gridDim.y != 1 && threadIdx.x == 0) {
dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
dst_meta[j_dst_unrolled] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
}
}
#else