Issues

Load quantized storage formats (ModelOpt NVFP4, MLX, compressed-tensors with activation scales, bitsandbytes, AutoRound, fbgemm FP8)

AshishKumar4/dew#33 · opened

on Oct 5, 2026

1

bf16 MoE layout parity: 5758x at layers_3 expert gate_proj on expert4 (cause not established)

AshishKumar4/dew#21 · opened

on Oct 2, 2026

2

SSD MoE streaming: investigate slow host mmap read and conversion

AshishKumar4/dew#20 · opened

on Oct 2, 2026

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

AshishKumar4/dew#19 · opened

on Sep 25, 2026

1

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

AshishKumar4/dew#18 · opened

on Sep 25, 2026

3

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

AshishKumar4/dew#17 · opened

on Sep 25, 2026

4

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

AshishKumar4/dew#16 · opened

on Sep 25, 2026

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

AshishKumar4/dew#15 · opened

on Sep 25, 2026

1

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

AshishKumar4/dew#14 · opened

on Sep 25, 2026

Measure SGLang at TP=2 and TP=4 through OpenAIRolloutServer against vLLM and Dew

AshishKumar4/dew#13 · opened

on Sep 25, 2026

Measure the NCCL weight push to vLLM at gpt-oss-20b scale

AshishKumar4/dew#12 · opened

on Sep 25, 2026

Check that a 4-process checkpoint resume is bit-exact against a 4-process continuous run

AshishKumar4/dew#11 · opened

on Sep 25, 2026

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

AshishKumar4/dew#10 · opened

on Sep 25, 2026

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

AshishKumar4/dew#9 · opened

on Sep 25, 2026

Measure training throughput per layout for every model, and state the rule for choosing one

AshishKumar4/dew#8 · opened

on Sep 25, 2026

Decide whether GPU training turns on XLA's latency-hiding scheduler by default

AshishKumar4/dew#7 · opened

on Sep 25, 2026

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

AshishKumar4/dew#6 · opened

on Sep 25, 2026

XLA SPMD halo convolution is wrong with an input-feature-split kernel (Conv keeps its kernel whole)

AshishKumar4/dew#4 · opened

on Sep 25, 2026

2

XLA:CPU scan over a chunked delta-rule layer returns process-dependent wrong values (KDA barrier)

AshishKumar4/dew#3 · opened

on Sep 25, 2026

2

XLA:TPU folds `i > 0` on a loop counter that starts at a constant 0 (UniPC works around it)

AshishKumar4/dew#1 · opened

on Sep 25, 2026

2

Navigation Menu

Repository navigation

Issues

Load quantized storage formats (ModelOpt NVFP4, MLX, compressed-tensors with activation scales, bitsandbytes, AutoRound, fbgemm FP8)

bf16 MoE layout parity: 5758x at layers_3 expert gate_proj on expert4 (cause not established)

SSD MoE streaming: investigate slow host mmap read and conversion

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

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

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

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

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

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

Measure SGLang at TP=2 and TP=4 through OpenAIRolloutServer against vLLM and Dew

Measure the NCCL weight push to vLLM at gpt-oss-20b scale

Check that a 4-process checkpoint resume is bit-exact against a 4-process continuous run

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

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

Measure training throughput per layout for every model, and state the rule for choosing one

Decide whether GPU training turns on XLA's latency-hiding scheduler by default

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

XLA SPMD halo convolution is wrong with an input-feature-split kernel (Conv keeps its kernel whole)

XLA:CPU scan over a chunked delta-rule layer returns process-dependent wrong values (KDA barrier)

XLA:TPU folds `i > 0` on a loop counter that starts at a constant 0 (UniPC works around it)