[Quantization][Humming] Support MXFP4 weight + block-FP8 activation for MoE (#51332)

Signed-off-by: Elvir Crncevic <elvircrn@gmail.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Signed-off-by: Elvir Crnčević <elvircrn@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Elvir Crnčević
2026-08-25 10:21:00 -04:00
committed by GitHub
co-authored by Claude Opus 4.8 Tyler Michael Smith mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 9c8e90eb26
commit ca29cd48ba
17 changed files with 1594 additions and 171 deletions
+10 -1
View File
@@ -611,7 +611,15 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
// SituGLU implementation used in Kimi models.
ops.def(
"situ_and_mul(Tensor! out, Tensor input, float beta=1.0, float "
"linear_beta=-1.0) -> ()");
"linear_beta=-1.0, Tensor? valid_rows=None) -> ()");
// Fused SituGLU activation + dynamic FP8 quantization for the Humming w2 path
// (writes the fp8 down input and its float32 scale). group_size=0 ->
// per-token scale [.., 1]; group_size=128 -> k-major block-FP8 scale [..,
// d/128].
ops.def(
"situ_and_mul_quant(Tensor! out, Tensor! scale, Tensor input, "
"float beta=1.0, float linear_beta=-1.0, int group_size=0, "
"Tensor? valid_rows=None, int topk=1) -> ()");
ops.def(
"masked_situ_and_mul(Tensor! out, Tensor input, Tensor "
"expert_num_tokens, float beta=1.0, float linear_beta=-1.0) -> ()");
@@ -830,6 +838,7 @@ STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, ops) {
ops.impl("fatrelu_and_mul", TORCH_BOX(&fatrelu_and_mul));
ops.impl("swigluoai_and_mul", TORCH_BOX(&swigluoai_and_mul));
ops.impl("situ_and_mul", TORCH_BOX(&situ_and_mul));
ops.impl("situ_and_mul_quant", TORCH_BOX(&situ_and_mul_quant));
ops.impl("masked_situ_and_mul", TORCH_BOX(&masked_situ_and_mul));
ops.impl("gelu_new", TORCH_BOX(&gelu_new));
ops.impl("gelu_fast", TORCH_BOX(&gelu_fast));