1 Release published by 1 person

• v1.1.0 FBGEMM_GPU v1.1.0 Release Notes

  published Jan 29, 2025

19 Pull requests opened by 8 people

• Fix handling of dynamic FP8 grouped gemm on Nvidia

  #3616 opened Jan 26, 2025

• Performance Optimization: Optimized TileShape Configuration for f8

  #3617 opened Jan 27, 2025

• Re-land D67407935 (Optimized backward pass for ROCm devices, pt 2)

  #3619 opened Jan 27, 2025

• Partial revert of D66986498

  #3620 opened Jan 27, 2025

• finish #1808 cherry-pick, adjust interface

  #3627 opened Jan 28, 2025

• Update bf16i4 gemm with new cutlass version

  #3630 opened Jan 29, 2025

• avoid using warning tensor in cpu tbe op

  #3631 opened Jan 29, 2025

• k_norm in rope for fp8 kv cache

  #3633 opened Jan 29, 2025

• Partial revert of D66986498 (Optimized backward pass for ROCm devices, pt 1), 2nd attempt

  #3637 opened Jan 29, 2025

• Adding Missing includes and explicitly declaring Tensor in aten namespace.

  #3638 opened Jan 30, 2025

• Fix zero_start_index_M argument for triton rowwise quantize

  #3639 opened Jan 30, 2025

• Re-organize SLL ops, pt 4

  #3644 opened Jan 30, 2025

• Adding couple more APIs to KVTensorWrapper to bring partiy with torch::Tensor

  #3645 opened Jan 31, 2025

• Re-organize SLL ops, pt 5

  #3646 opened Jan 31, 2025

• Re-organize SLL ops, pt 6

  #3647 opened Jan 31, 2025

• Add preprocess stage to quantize bench operators

  #3648 opened Jan 31, 2025

• Re-organize SLL ops, pt 7

  #3650 opened Jan 31, 2025

• Add tracing option to quantize bench

  #3651 opened Jan 31, 2025

• Improve FP8 grouped GEMM perf via tileshape and cooperative

  #3653 opened Feb 2, 2025

26 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Use int64_t for buffer indices to avoid overflow

  #1896 commented on Jan 27, 2025 • 0 new comments

• Fixing type comparing errors from Lint

  #1911 commented on Jan 27, 2025 • 0 new comments

• Debug test_jagged_dense_bmm CI errors

  #1928 commented on Jan 27, 2025 • 0 new comments

• Print CMake commands

  #1949 commented on Jan 27, 2025 • 0 new comments

• Use Nova workflow to host all published wheel files at PyTorch site

  #1958 commented on Jan 27, 2025 • 0 new comments

• Revert diff that fails test

  #1995 commented on Jan 27, 2025 • 0 new comments

• Fix hypothesis version

  #1996 commented on Jan 27, 2025 • 0 new comments

• Revert skipping test_pack_segments

  #2037 commented on Jan 27, 2025 • 0 new comments

• Add device parameter to DenseTableBatchedEmbeddingBagsCodegen

  #2054 commented on Jan 27, 2025 • 0 new comments

• Unify TBE API for key FBGEMM operations (Frontend)

  #2081 commented on Jan 27, 2025 • 0 new comments

• Skip InputCombineTest for ROCm on OSS CI

  #2229 commented on Jan 27, 2025 • 0 new comments

• Add header to fix PRIu64 error

  #2234 commented on Jan 27, 2025 • 0 new comments

• ensemble_rowwise_adagrad

  #2871 commented on Jan 27, 2025 • 0 new comments

• Remove SSD from OSS

  #3043 commented on Jan 27, 2025 • 0 new comments

• build cuda-only

  #3061 commented on Jan 27, 2025 • 0 new comments

• Test build time for sm 8.0 only

  #3068 commented on Jan 27, 2025 • 0 new comments

• Fix FBGEMM CI strict aliasing and unused var errors

  #3083 commented on Jan 27, 2025 • 0 new comments

• Add NEON and SVE implementations for Float16 conversions

  #3424 commented on Jan 30, 2025 • 0 new comments

• Remove numpy requirement

  #3470 commented on Jan 27, 2025 • 0 new comments

• sparse_ema_options (fbgemm)

  #3473 commented on Jan 27, 2025 • 0 new comments

• Support FP8 grouped GEMM with rowwise scailing

  #3560 commented on Feb 2, 2025 • 0 new comments

• Refactor FP8 grouped GEMM with dynamic and static versions

  #3561 commented on Feb 2, 2025 • 0 new comments

• Unifying TBE API using List (Backend)

  #3563 commented on Jan 27, 2025 • 0 new comments

• AdagradW

  #3605 commented on Feb 2, 2025 • 0 new comments

• Port oss f16_fast_gemv into fbcode

  #3610 commented on Feb 1, 2025 • 0 new comments

• Updating split_table_batched_embeddings_ops_training.py

  #3613 commented on Jan 31, 2025 • 0 new comments