Skip to content

[Train] Add DeepSeek Engram#1107

Open
lxd-cumt wants to merge 15 commits intoflagos-ai:mainfrom
lxd-cumt:ds_rep
Open

[Train] Add DeepSeek Engram#1107
lxd-cumt wants to merge 15 commits intoflagos-ai:mainfrom
lxd-cumt:ds_rep

Conversation

@lxd-cumt
Copy link
Collaborator

@lxd-cumt lxd-cumt commented Feb 4, 2026

Add deepseek engram, ref: DeepSeek Engram Paper, DeepSeek Engram Github

  • Support tensor parallel, pipeline parallel, sequence parallel, distributed data parallel
  • Support NgramHash Caching
  • End to end training support
  • CI/CD Tests
  • CKPT Conversion: FlagScale to HuggingFace

TODO:

  • Engram Embedding Offload
  • Engram Prefetch, attn/mlp computation and memory access overlapping
  • FlagOS support, based on Megatron-LM-FL and TransformerEngine-FL

update ds_v3 yamls

support deepseek engram, first version

add engram yamls

fix import transformer_block errors

fix dict

debug

fix engram config

fix engram config

seperate engram_transformer_layer and orig_transformer_layer

debug

fix layer_ids offset or engram and mcore

fix device error in compressed_tokenizer

update tokenizer path

fix

fix device error

debug pritn

fake hyper-connections, mhc to be supported

debug

disable sequence parallel, to be supported

disable tp, moe force sp with tp

debug

debug, reset print

update output dir name

debug multi-head-embedding

debug

support tp/sp embedding

update

tp/sp support

update yamls

enable tp/sp

debug print

add debug print

debug pp

debug layer id

debug pp

debug

update engram layer for pp

fix

debug offset and num_layers

fix layer ids offsets

debug pp

debug mtp

enable mtp

debug mtp

update mtp

fix

udpate pp size

add comment

debug pp

update pp test

modify get_batch for pp

fix

polish print

update engram yaml config

polish engram config names

fix

fix

debug conv

debug print

rewrite engram_model init

update

update engram model init

debug

add engram arguments

update yaml

update

polish

add ngram_hash cache

add nvtx profile

support nsys profile

from numpy to torch

polish print

debug print

update

polish print

update nvtx profile

opt nvtx

debug print

debug memcpy

opt nvtx

hash prefix

fix

fix

polish print

polish

unset yamls

unset nsys profile

remove run.sh
@lxd-cumt lxd-cumt force-pushed the ds_rep branch 2 times, most recently from a7b60a0 to 53c575f Compare February 11, 2026 02:16
disable bias linear

exclude engram hf models in pre-commit

support qwen3-engram ckpt conversion

update run.sh

fix

fix

fix

hack for tokenizer path

hack for tokenizer

fix

fix

fix

fix

fix

fix

fix

fix

fix

fix tokenizer path

fix

fix
update yamls

fix ssh port

fix

update deepseek yamls

polish
fix ruff check of cicd

rename

fix tp/pp input_ids transfer

fix

fix

unset path

update golden values

modify pre-commit for ci debug

fix

format

format
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant