Skip to content
Open
Show file tree
Hide file tree
Changes from 33 commits
Commits
Show all changes
36 commits
Select commit Hold shift + click to select a range
80e2d26
initial commit
clairesonglee Dec 10, 2025
d23d79f
set self.lr_warmup_steps < self.lr_decay_steps
clairesonglee Dec 16, 2025
3381850
unwrap model to remove loss_mask parameter
clairesonglee Dec 18, 2025
277f3e1
add zebra-llama (hybrid mla mamba model) support
Mingyuyang-1 Jan 13, 2026
11b22c6
add Zebra-Llama 3B configurations
Mingyuyang-1 Jan 22, 2026
1dec95e
add Zebra-Llama 1B configs and remove unused configs
Mingyuyang-1 Jan 23, 2026
34508db
remove unused configs
Mingyuyang-1 Jan 23, 2026
2f3ab49
Set submodule mamba to track enable-primus-hybrid-models branch
clairesonglee Jan 29, 2026
159e441
set moe_layer_freq default value of 1
clairesonglee Feb 3, 2026
f798e77
set final_logit_softcapping and router_logit_softcapping to null
clairesonglee Feb 3, 2026
d7f4faf
use mamba builder
clairesonglee Feb 4, 2026
4e4c445
Merge branch 'main' into clairlee/dev/hybrid
clairesonglee Feb 4, 2026
d5bdbb8
adjust zebra-llama architecture and training
Mingyuyang-1 Feb 5, 2026
ee92bb2
add gdn support
Mingyuyang-1 Mar 11, 2026
c8d32f3
support FSDP2 for hybrid model
Mingyuyang-1 Mar 12, 2026
f0f306a
add support to KDA layers
Mingyuyang-1 Mar 12, 2026
2b196f0
added patch.sh to fix fla autotune kernel bug for mi300.
vanshbhatia-amd Mar 19, 2026
1dc9a0e
working on matching h100 and primus losses
vanshbhatia-amd Apr 21, 2026
1a43061
training matches fla for gdn
vanshbhatia-amd Apr 24, 2026
550ef6b
Add GDN 300M training config and tools
vanshbhatia-amd May 6, 2026
c3fe815
Merge main into vanbhati/kda-optimized-training-patch
vanshbhatia-amd May 6, 2026
898a14c
adding patch
vanshbhatia-amd May 7, 2026
21c094d
Add bash-docker script and GDN training documentation
vanshbhatia-amd May 19, 2026
5638c01
Add KDA training documentation and configurations for FLA parity
vanshbhatia-amd May 20, 2026
f3280da
Add KDA training documentation and configurations for FLA parity
vanshbhatia-amd May 25, 2026
fdab2b3
Update .gitignore and remove all_changes.md
vanshbhatia-amd Jun 8, 2026
f6cef9b
Enhance GDN and KDA documentation for FLA parity
vanshbhatia-amd Jun 8, 2026
91227f8
enable mock data and training from no chkpt
clairesonglee Apr 22, 2026
256086b
fix: remove moe_layer from MambaStackSubmodules in hybrid_stack_spec
clairesonglee Jun 8, 2026
904a5d0
switch all zebra_llama configs to mock_data and remove user-specific …
clairesonglee Jun 8, 2026
eb47007
fix: disable overlap_grad_reduce for hybrid KDA/GDN 1B configs
clairesonglee Jun 8, 2026
fe56c25
tune zebra_llama configs: set train_iters=50, fix batch sizes and DDP…
clairesonglee Jun 8, 2026
8424d25
experimental gdn changes to avoid bf16 downcast
clairesonglee Jul 15, 2026
7ba8dc5
Merge remote-tracking branch 'origin/main' into clairlee/kda-optimize…
clairesonglee Jul 15, 2026
cb6fa5f
fix: pre-compute GDN gate in fp32 to avoid FLA Triton NaN on ROCm
clairesonglee Jul 15, 2026
733860c
Merge branch 'main' into clairlee/kda-optimized-training-patch
clairesonglee Jul 20, 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
22 changes: 22 additions & 0 deletions .gitignore
Original file line number Diff line number Diff line change
Expand Up @@ -12,3 +12,25 @@ output
experiment
data
pp_simulation_result

# Scratch run artefacts (repo root only; archived results under experiments/results/archive/ ARE tracked)
/primus_*.log
/primus_*.launcher.out
/primus_perf_*.summary.txt
/experiments/results/*.log
/experiments/results/*.launcher.out
/experiments/results/*.driver.out
/train_gdn.log
/run_pure_1B_eval.log
eval_results/

# FLA-parity evaluation harness — local working copy only (configs/launchers/
# logs/summaries used to iterate on the 300M & 1B comparison runs; not part
# of the Primus public surface, kept on disk for follow-up work).
/eval/

# IDE / editor workspace files (not part of source)
Primus.code-workspace
**/Primus.code-workspace
*.code-workspace
.cursor/
4 changes: 4 additions & 0 deletions .gitmodules
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,10 @@
[submodule "third_party/Megatron-Bridge"]
path = third_party/Megatron-Bridge
url = https://github.com/NVIDIA-NeMo/Megatron-Bridge.git
[submodule "third_party/mamba"]
path = third_party/mamba
url = https://github.com/AndreasKaratzas/mamba.git
branch = enable-primus-hybrid-models
[submodule "third_party/HummingbirdXT"]
path = third_party/HummingbirdXT
url = https://github.com/AMD-AGI/HummingbirdXT.git
309 changes: 309 additions & 0 deletions GDN_FLA_PARITY.md

Large diffs are not rendered by default.

276 changes: 276 additions & 0 deletions KDA_FLA_PARITY.md

Large diffs are not rendered by default.

7 changes: 7 additions & 0 deletions bash-docker.sh
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
docker run -it \
--device /dev/dri --device /dev/kfd \
--device=/dev/infiniband --network host --ipc host \
--group-add video --cap-add SYS_PTRACE \
--security-opt seccomp=unconfined --privileged \
-v $HOME:$HOME -v $(pwd):$(pwd) -w $(pwd) --shm-size 64G --name primus_hybrid_new \
rocm/primus:v26.2
Loading