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

CUDA: dedup MoE gate/up activation quantization (#25441)

Browse Source 
* CUDA: dedup MoE gate/up activation quantization (fp4)

For MoE gate/up projections the src1 activation is broadcast across the
routed experts (ne11 == 1), so ids_src1 maps every one of a token's
n_expert_used slots to the same physical row. The MMQ path therefore
re-quantized each token's activation n_expert_used times.

For fp4 (NVFP4/MXFP4) src0, quantize each unique token row once instead of
once per expert. For NVFP4 a single quantize+scatter kernel
(quantize_scatter_mmq_nvfp4) quantizes each token once and writes the
resulting block_fp4_mmq straight to all n_expert_used slots, using an
inverse token->compact-row map (build_tok2c). MXFP4, and
GGML_CUDA_MOE_QUANT_GATHER=1, use a two-kernel variant: quantize unique
rows then gather into the expert-sorted layout (gather_mmq_fp4_blocks).
Both are bit-identical to the previous gather-then-quantize path (identical
source data, deterministic per-block quantization), verified by
test-backend-ops MUL_MAT_ID (type_a=nvfp4, broadcast b=1; 790/790 for the
default, gather, and per-expert paths) and by coherent end-to-end
generation. Set GGML_CUDA_NO_MOE_QUANT_DEDUP=1 to force the original
per-expert path.

Same-binary A/B on RTX 5090 (sm_120), Qwen3.6-35B-A3B-NVFP4 prefill @8192
(nsys, graphs-off; the unchanged mul_mat_q GEMM confirms stable clocks):
activation-quant GPU-busy drops 61% (78.2 -> 30.4 ms) with the fused
quantize+scatter, vs 33% (78.2 -> 52.8 ms) for the two-kernel gather. The
fused path avoids materializing and re-reading the 8x compact buffer,
writing the expert copies directly from registers.

* CUDA: bounds-check token ids in build_tok2c_kernel

Guard against malformed ids_src1: skip out-of-range token ids (t < 0 or
t >= n_tokens) and drop entries beyond n_expert_used per token instead of
writing past the token's tok2c region. No behavior change for valid MoE
routing data; test-backend-ops MUL_MAT_ID 790/790.

* Refactor the code based on review comments

- Removed previously added kernels that were not necessary anymore\
- Added an inverse mapping from (token, slot) to compact row. Each token is quantized once and scattered to its compact rows.

* Adding q8_1 support for dedup and addressing review comments

* Add pragma unrolls

* Remove redundant cudaMemsetAsync call

* Removing follow up redundancies

---------

Co-authored-by: praneshgo <227579474+praneshgo@users.noreply.github.com>
This commit is contained in:
Pranesh Gonegandla
2026-07-16 09:02:25 +02:00
committed by GitHub
co-authored by praneshgo
parent a320cbfcb7
commit 5839ba3524
6 changed files with 255 additions and 116 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/mmid.cu
+18 -13
View File
@@ -27,7 +27,7 @@ template <int n_expert_used_template>
__launch_bounds__(ggml_cuda_get_physical_warp_size(), 1)
static __global__ void mm_ids_helper(
const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1) {
const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1, const bool write_inverse) {
constexpr int warp_size = ggml_cuda_get_physical_warp_size();
const int n_expert_used = n_expert_used_template == 0 ? n_expert_used_var : n_expert_used_template;
const int expert = blockIdx.x;
@@ -98,8 +98,13 @@ static __global__ void mm_ids_helper(
const mm_ids_helper_store store_it = store[itc];
const int it = store_it.it();
const int iex_used = store_it.iex_used();
ids_src1[nex_prev + itc] = it*sis1 + iex_used % nchannels_y;
ids_dst [nex_prev + itc] = it*n_expert_used + iex_used;
ids_dst[nex_prev + itc] = it*n_expert_used + iex_used;
// ids_src1 holds the forward map, or the inverse map (token slot -> compact row) for quant dedup
if (write_inverse) {
ids_src1[it*n_expert_used + iex_used] = nex_prev + itc;
} else {
ids_src1[nex_prev + itc] = it*sis1 + iex_used % nchannels_y;
}
}
if (threadIdx.x != 0) {
@@ -118,7 +123,7 @@ static __global__ void mm_ids_helper(
template <int n_expert_used_template>
static void launch_mm_ids_helper(
const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
const int n_experts, const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1, cudaStream_t stream) {
const int n_experts, const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1, const bool write_inverse, cudaStream_t stream) {
GGML_ASSERT(n_tokens < (1 << 22) && "too few bits in mm_ids_helper_store");
GGML_ASSERT(n_expert_used_var < (1 << 10) && "too few bits in mm_ids_helper_store");
@@ -132,33 +137,33 @@ static void launch_mm_ids_helper(
const size_t nbytes_shared = n_tokens*sizeof(mm_ids_helper_store);
GGML_ASSERT(nbytes_shared <= smpbo);
mm_ids_helper<n_expert_used_template><<<num_blocks, block_size, nbytes_shared, stream>>>
(ids, ids_src1, ids_dst, expert_bounds, n_tokens, n_expert_used_var, nchannels_y, si1, sis1);
(ids, ids_src1, ids_dst, expert_bounds, n_tokens, n_expert_used_var, nchannels_y, si1, sis1, write_inverse);
}
void ggml_cuda_launch_mm_ids_helper(
const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
const int n_experts, const int n_tokens, const int n_expert_used, const int nchannels_y, const int si1, const int sis1, cudaStream_t stream) {
const int n_experts, const int n_tokens, const int n_expert_used, const int nchannels_y, const int si1, const int sis1, const bool write_inverse, cudaStream_t stream) {
switch (n_expert_used) {
case 2:
launch_mm_ids_helper< 2>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
launch_mm_ids_helper< 2>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
break;
case 4:
launch_mm_ids_helper< 4>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
launch_mm_ids_helper< 4>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
break;
case 6:
launch_mm_ids_helper< 6>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
launch_mm_ids_helper< 6>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
break;
case 8:
launch_mm_ids_helper< 8>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
launch_mm_ids_helper< 8>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
break;
case 16:
launch_mm_ids_helper<16>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
launch_mm_ids_helper<16>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
break;
case 32:
launch_mm_ids_helper<32>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
launch_mm_ids_helper<32>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
break;
default:
launch_mm_ids_helper< 0>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
launch_mm_ids_helper< 0>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
break;
}
}