Skip to content

Enable NVFP4 grouped MLP GLU RHT amax path #4

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

Open
sraman-rgb wants to merge 26 commits into
base: main
Choose a base branch
from
Open

Enable NVFP4 grouped MLP GLU RHT amax path #4

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
26 commits
Select commit Hold shift + click to select a range
3f1d889
Fix WHEEL Tag mismatch in transformer-engine-cu12 wheels (#2928)
eyupcanakman Jun 2, 2026
b24049b
[Fix] Fix CUTLASS grouped GEMM segfault for empty groups (#3067)
Baibaifan Jun 2, 2026
9028a39
Enable NVFP4 grouped MLP SReLU fusion (#3072)
sraman-rgb Jun 2, 2026
e8102e6
Enable NVFP4 grouped MLP cuDNN wgrad (#3071)
sraman-rgb Jun 2, 2026
f5e500b
Bitmap topk (#3009)
tdophung Jun 2, 2026
c1e827f
[PyTorch] Refactor function to prepare pointers for grouped MLP discr…
timmoon10 Jun 3, 2026
3bca938
increasing precision tolerance (#3060)
francesco-bertolotti Jun 3, 2026
54e5bfc
[JAX] Fallback to old triton ffi for autotuned kernels (#3077)
jberchtold-nvidia Jun 3, 2026
b38c16b
Optimize grouped split metadata preparation (#3075)
zhongbozhu Jun 3, 2026
815bf36
[Common] Comm+GEMM overlap API updated to support cuBlasMp backend (i…
denera Jun 4, 2026
86d4e15
[PyT] Reduce test sizes in fused attn fp8 vs fp16 to avoid OOM (#3020)
vedaanta Jun 4, 2026
5535b09
[PyTorch] Fix FlashAttention 2 head_dim > 192 on sm103 and other arch…
pedramr Jun 4, 2026
64311fe
Add MXFP8 attention unit test with linear and rope layers (#3033)
layalir Jun 4, 2026
1b12177
[PyTorch] Expose interleave and de-interleave function for GLU tensor…
ksivaman Jun 4, 2026
abdb406
[JAX] Support for cuDNN-backed flex attention (#2985)
vcherepanov-nv Jun 4, 2026
97a9bfe
[PyTorch] Support for cuDNN-backed flex attention (#2984)
vcherepanov-nv Jun 4, 2026
f458abe
[PyTorch] Python DType enum (#3039)
vthumbe1503 Jun 4, 2026
79f163a
Enable NVFP4 grouped MLP GLU RHT amax path
sraman-rgb May 30, 2026
f7aba76
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Jun 1, 2026
f3b03d7
Address NVFP4 GLU RHT amax review comments
sraman-rgb Jun 1, 2026
eeb65bf
Deduplicate grouped NVFP4 quantize helper
sraman-rgb Jun 2, 2026
d4bc09b
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Jun 2, 2026
98e3210
Route precomputed amax through NVFP4 quantize
sraman-rgb Jun 2, 2026
56a8e50
Handle empty NVFP4 precomputed amax reduction
sraman-rgb Jun 2, 2026
d3533df
Fix NVFP4 amax quantize binding signature
sraman-rgb Jun 4, 2026
8eef24c
Merge branch 'main' into nvfp4-grouped-mlp-glu-rht-amax
sraman-rgb Jun 4, 2026
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
5 changes: 2 additions & 3 deletions benchmarks/benchmark_rht_cast.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,6 @@
import torch.utils.benchmark as benchmark

import transformer_engine.pytorch as te
import transformer_engine_torch as tex
import transformer_engine.pytorch.cpp_extensions as ext

from transformer_engine.pytorch.tensor.nvfp4_tensor import NVFP4Quantizer
Expand All @@ -17,7 +16,7 @@
permute_scale = False

TORCH_TO_TE_FLOAT_MAP = {
torch.bfloat16: tex.DType.kBFloat16,
torch.bfloat16: te.DType.kBFloat16,
}


Expand All @@ -31,7 +30,7 @@ def run_kernel(shape, stochastic_rounding: bool, input_dtype=torch.bfloat16):

# Quantize
nvfp4_quantizer = NVFP4Quantizer(
fp4_dtype=tex.DType.kFloat4E2M1,
fp4_dtype=te.DType.kFloat4E2M1,
rowwise=True,
columnwise=True,
with_amax_reduction=False,
Expand Down
Loading
Loading