Skip to content

[Speculative Decoding][MTP] Support static CacheKV C8 quantization and optimize memory usage #5155

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
freeliuzc merged 2 commits into from
Nov 21, 2025
Merged

[Speculative Decoding][MTP] Support static CacheKV C8 quantization and optimize memory usage #5155

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d257af9
support static cachekv c8 quantization in mtp mode
freeliuzc Nov 20, 2025
8a6b51d
optimize memory allocation
freeliuzc Nov 20, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
42 changes: 25 additions & 17 deletions custom_ops/gpu_ops/append_attn/speculate_write_cache_with_rope_impl.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -602,7 +602,8 @@ template <typename T,
int VecSize = 4,
int RoundType = 0,
int HeadDim = 128,
bool IsFP8 = false>
bool IsFP8 = false,
bool IsDynamic = true>
__global__ void append_speculate_cache_fp8_rope_qk_norm_dynamic_kernel(
const T* __restrict__ quant_qkv, // [num_head, num_heads + 2 *
// gqa_group_size, head_size]
Expand Down Expand Up @@ -662,8 +663,6 @@ __global__ void append_speculate_cache_fp8_rope_qk_norm_dynamic_kernel(
(head_idx - num_heads) % gqa_group_size * block_size +
block_offset;
}
T* cache_k_scale_now = cache_k_scale + cache_offset;
T* cache_v_scale_now = cache_v_scale + cache_offset;

float thread_m2 = 0.0f;
float warp_m2 = 0.0f;
Expand Down Expand Up @@ -811,25 +810,34 @@ __global__ void append_speculate_cache_fp8_rope_qk_norm_dynamic_kernel(
}
}
// reduce max, 1 head per warp
T local_max = -INFINITY;
if constexpr (IsDynamic) {
T local_max = -INFINITY;
#pragma unroll
for (int i = 0; i < HALF_K_VEC_SIZE; i++) {
local_max = __hmax(local_max, __habs(bias_vec1[i]));
local_max = __hmax(local_max, __habs(bias_vec2[i]));
}
for (int i = 0; i < HALF_K_VEC_SIZE; i++) {
local_max = __hmax(local_max, __habs(bias_vec1[i]));
local_max = __hmax(local_max, __habs(bias_vec2[i]));
}
#pragma unroll
for (int m_offset = 16; m_offset > 0; m_offset /= 2) {
local_max =
__hmax(local_max, __shfl_xor_sync(0xffffffff, local_max, m_offset));
}

scale = __hdiv(448, local_max);
for (int m_offset = 16; m_offset > 0; m_offset /= 2) {
local_max =
__hmax(local_max, __shfl_xor_sync(0xffffffff, local_max, m_offset));
}

if (lane_id == 0) {
scale = __hdiv(448, local_max);
T* cache_k_scale_now = cache_k_scale + cache_offset;
T* cache_v_scale_now = cache_v_scale + cache_offset;
if (lane_id == 0) {
if (head_idx < num_heads + gqa_group_size) {
cache_k_scale_now[0] = __hdiv(1, scale);
} else {
cache_v_scale_now[0] = __hdiv(1, scale);
}
}
} else {
if (head_idx < num_heads + gqa_group_size) {
cache_k_scale_now[0] = __hdiv(1, scale);
scale = __ldg(&cache_k_scale[kv_head_idx]);
} else {
cache_v_scale_now[0] = __hdiv(1, scale);
scale = __ldg(&cache_v_scale[kv_head_idx]);
}
}

Expand Down
Loading
Loading