Skip to content

[None][perf] Staircase: Rubin GLM C1 Megakernel #21590

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
dongfengy merged 5 commits into from
Nov 8, 2026
+15,904 −1
Merged

[None][perf] Staircase: Rubin GLM C1 Megakernel #21590

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
179f949
[None][perf] Add Rubin C1 GLM megakernel to the staircase
dongfengy Nov 4, 2026
20fb70c
[None][fix] Reserve speculative headroom in the Rubin megakernel
dongfengy Nov 4, 2026
972b2a6
[None][fix] Keep the megakernel setup query on its native implementation
dongfengy Nov 7, 2026
71eaa66
[None][fix] Honor the CUDA tensor-map type alignment
dongfengy Nov 7, 2026
42bbd16
[None][fix] Satisfy GLM megakernel native operator contracts
dongfengy Nov 7, 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
33 changes: 33 additions & 0 deletions tensorrt_llm/_torch/_experimental/modeling_v2/models/glm_moe_dsa/csrc/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,39 @@ add_modeling_v2_target(
ARCHS 100f
SOURCES *.cu *.cpp)

# Preserve the C1 megakernel specialization and native Rubin code generation.
set(mega_source
"${CMAKE_CURRENT_SOURCE_DIR}/../glm_5_2_nvfp4__sm_107__low_latency/csrc/glmMegakernel.cu"
)
set_property(
SOURCE "${mega_source}"
APPEND
PROPERTY COMPILE_DEFINITIONS
GLM52_FC2_SPARSE_WIDE=1
GLM52_OWNER_FC1_OVERLAP=1
GLM52_ROUTER_LG=2
GLM52_CAND_FILTER_ROLL=0
GLM52_SEL_BATCH_MMA=1
GLM52_LEAN=1
GLM52_FLAGS_MODE=5
GLM52_IO_MODE=3
GLM52_STAMPS_MODE=1
GLM52_SETUP_MODE=1
GLM52_LISTS_MODE=2
GLM52_TOPK_UNION=1
GLM52_TOPK_RSH=1
GLM52_STATIC_FLAGS=2752ll
GLM52_SMEM_MAIN=200704)
set_property(
SOURCE "${mega_source}" APPEND
PROPERTY COMPILE_OPTIONS --maxrregcount=200 --ftz=false --prec-div=true
--prec-sqrt=true)
if(107 IN_LIST CMAKE_CUDA_ARCHITECTURES_ORIG)
set_property(
SOURCE "${mega_source}" APPEND
PROPERTY COMPILE_OPTIONS --generate-code=arch=compute_107a,code=sm_107a)
endif()

add_modeling_v2_target(
glm_moe_dsa glm_5_2_nvfp4__sm_107__low_latency
ARCHS 100f
Expand Down
Loading
Loading