AshishKumar4/dew

Load hybrid and state-space decoders (nemotron_h, lfm2, falcon_h1, zamba2, jamba) 27 days ago

enhancement help wanted

0

A nice ML framework based on Jax/Flax. Based on https://github.com/AshishKumar4/FlaxDiff

Python

#diffusion-models#distributed-training#flax#flax-linen#fsdp#jax#jepa#llm-training#llm-training-from-scratch#ml-library

Load multimodal wrappers around registered text families (qwen3_5_moe, gemma3/gemma4, qwen3_vl_moe, mistral3) 27 days ago

enhancement help wanted

AshishKumar4/dew

0

A nice ML framework based on Jax/Flax. Based on https://github.com/AshishKumar4/FlaxDiff

Python

#diffusion-models#distributed-training#flax#flax-linen#fsdp#jax#jepa#llm-training#llm-training-from-scratch#ml-library

Load classic GPT-style decoders (gpt2, opt, gpt_neox, bloom, falcon, phi, gptj, bigcode, starcoder2, stablelm, cohere) 27 days ago

enhancement help wanted

AshishKumar4/dew

0

A nice ML framework based on Jax/Flax. Based on https://github.com/AshishKumar4/FlaxDiff

Python

#diffusion-models#distributed-training#flax#flax-linen#fsdp#jax#jepa#llm-training#llm-training-from-scratch#ml-library

Run the regression tests that need multi-GPU pools, multi-chip TPUs or a specific GPU about 1 month ago

help wanted needs hardware

AshishKumar4/dew

0

A nice ML framework based on Jax/Flax. Based on https://github.com/AshishKumar4/FlaxDiff

Python

#diffusion-models#distributed-training#flax#flax-linen#fsdp#jax#jepa#llm-training#llm-training-from-scratch#ml-library

Mosaic GPU grouped matmul for sm90+ (the Pallas Triton backend is deprecated in jax 0.11.2) about 1 month ago

enhancement help wanted needs hardware

AshishKumar4/dew

0

A nice ML framework based on Jax/Flax. Based on https://github.com/AshishKumar4/FlaxDiff

Python

#diffusion-models#distributed-training#flax#flax-linen#fsdp#jax#jepa#llm-training#llm-training-from-scratch#ml-library

First two-host runs: hybrid sharding, the sequence exchange and expert parallelism across nodes about 1 month ago

enhancement help wanted needs hardware

AshishKumar4/dew

0

A nice ML framework based on Jax/Flax. Based on https://github.com/AshishKumar4/FlaxDiff

Python

#diffusion-models#distributed-training#flax#flax-linen#fsdp#jax#jepa#llm-training#llm-training-from-scratch#ml-library

Measure SGLang at TP=2 and TP=4 through OpenAIRolloutServer against vLLM and Dew about 1 month ago

enhancement help wanted needs hardware

AshishKumar4/dew

0

A nice ML framework based on Jax/Flax. Based on https://github.com/AshishKumar4/FlaxDiff

Python

#diffusion-models#distributed-training#flax#flax-linen#fsdp#jax#jepa#llm-training#llm-training-from-scratch#ml-library

Measure the NCCL weight push to vLLM at gpt-oss-20b scale about 1 month ago

enhancement help wanted needs hardware

AshishKumar4/dew

0

A nice ML framework based on Jax/Flax. Based on https://github.com/AshishKumar4/FlaxDiff

Python

#diffusion-models#distributed-training#flax#flax-linen#fsdp#jax#jepa#llm-training#llm-training-from-scratch#ml-library

Check that a 4-process checkpoint resume is bit-exact against a 4-process continuous run about 1 month ago

enhancement help wanted needs hardware

AshishKumar4/dew

0

A nice ML framework based on Jax/Flax. Based on https://github.com/AshishKumar4/FlaxDiff

Python

#diffusion-models#distributed-training#flax#flax-linen#fsdp#jax#jepa#llm-training#llm-training-from-scratch#ml-library

Re-check gpt-oss-20b serving on expert=4 now that a forward reads expert kernels as stored about 1 month ago

bug help wanted needs hardware

AshishKumar4/dew

0

A nice ML framework based on Jax/Flax. Based on https://github.com/AshishKumar4/FlaxDiff

Python

#diffusion-models#distributed-training#flax#flax-linen#fsdp#jax#jepa#llm-training#llm-training-from-scratch#ml-library

Serving on data=4 as one program reaches 0.71x of four one-GPU servers (host-bound) about 1 month ago

bug help wanted needs hardware

AshishKumar4/dew

0

A nice ML framework based on Jax/Flax. Based on https://github.com/AshishKumar4/FlaxDiff

Python

#diffusion-models#distributed-training#flax#flax-linen#fsdp#jax#jepa#llm-training#llm-training-from-scratch#ml-library

Measure training throughput per layout for every model, and state the rule for choosing one about 1 month ago

enhancement help wanted needs hardware

AshishKumar4/dew

0

A nice ML framework based on Jax/Flax. Based on https://github.com/AshishKumar4/FlaxDiff

Python

#diffusion-models#distributed-training#flax#flax-linen#fsdp#jax#jepa#llm-training#llm-training-from-scratch#ml-library

Decide whether GPU training turns on XLA's latency-hiding scheduler by default about 1 month ago

enhancement help wanted needs hardware

AshishKumar4/dew

0

A nice ML framework based on Jax/Flax. Based on https://github.com/AshishKumar4/FlaxDiff

Python

#diffusion-models#distributed-training#flax#flax-linen#fsdp#jax#jepa#llm-training#llm-training-from-scratch#ml-library

4-GPU training throughput on main vs PyTorch DDP and FSDP2 (dense, MoE, DiT) about 1 month ago

enhancement help wanted needs hardware

AshishKumar4/dew

0

A nice ML framework based on Jax/Flax. Based on https://github.com/AshishKumar4/FlaxDiff

Python

#diffusion-models#distributed-training#flax#flax-linen#fsdp#jax#jepa#llm-training#llm-training-from-scratch#ml-library

Open Issues Need Help