AI-Hypercomputer maxtext issues

AI-Hypercomputer / maxtext

A simple, performant and scalable Jax LLM!

Apache License 2.0

1.54k stars 295 forks source link

issues

Newest

Newest Most commented Recently updated Oldest Least commented Least recently updated

Make tolerance configurable

#1058 Doris26 opened 16 hours ago
1
Set block_q_dq and block_kv_dq to None if use_fused_bwd_kernel is enabled

#1057 raymondzouu opened 17 hours ago
0
More flexible offline benchmarking.

#1056 patemotter opened 18 hours ago
0
Adds 405B script for GPU for consistency.

#1055 wang2yn84 closed 15 hours ago
0
Add 32x8 custom split

#1054 RissyRan closed 13 hours ago
1
set tolerance a configurable param with default value 0.1

#1053 Doris26 closed 21 hours ago
1
Set Orbax logger to False.

#1052 abhinavclemson closed 1 day ago
1
Adding support for mixed precision drq.

#1051 singh-mitali closed 1 day ago
0
Updating Jax Stable Stack BaseImage path

#1050 parambole closed 2 days ago
0
Merge the disaggregation prototyping code

#1049 zhihaoshan-google opened 2 days ago
1
Add checkpoint topology discovery for the replicator service

#1048 xuefgu opened 3 days ago
0
Flash attention - head_dim 64

#1047 peregilk opened 4 days ago
0
Enable orbax checkpoint cloud logger as default.

#1046 abhinavclemson closed 3 days ago
0
[DO NOT MERGE] verify fix

#1045 RissyRan opened 6 days ago
0
Config for Goodput with Pathways

#1044 dipannita08 closed 3 days ago
0
[Do Not Merge] v6e scale testing goodput

#1043 SujeethJinesh opened 6 days ago
0
Add Llama2-70b sparsecore collective model to trillium configs

#1042 Obliviour opened 6 days ago
0
Add missing period.

#1041 gobbleturk closed 6 days ago
0
Enable pathways workloads for v6e benchmarks

#1040 sadikneipp opened 1 week ago
0
Scan and remat only the outer pipeline iteration loop over microbatches

#1039 gobbleturk closed 1 day ago
0
Adding new JSS builds for verification

#1038 parambole closed 6 days ago
0
Add custom 4x64 mesh for single slice

#1037 raymondzouu closed 1 day ago
0
Remove aqt einsum for dropping

#1036 mailvijayasingh opened 1 week ago
1
Update sharding annotation for dropping

#1035 RissyRan closed 1 week ago
0
Enable orbax cloud logger by default.

#1034 abhinavclemson opened 1 week ago
0
Use python3 and fix bash syntax

#1033 guptaaka closed 1 week ago
0
ray fault tolerance

#1032 keshavb96 opened 1 week ago
0
[Inference Perf] Add autotuned xla flags to improve latency for v6e

#1031 wyzhang closed 1 week ago
0
Adding support for building Maxtext images with nightly version of JSS for GPUs

#1030 parambole closed 1 week ago
0
[DO NOT MERGE] Ranran hide a2a

#1029 RissyRan opened 1 week ago
4
Support for safetensor checkpoints

#1028 richjames0 opened 1 week ago
0
Support for safetensor checkpoints

#1027 richjames0 closed 1 week ago
0
[MoE] fix typo and add normalization for top_k_weights

#1026 ZhiyuLi-goog opened 1 week ago
7
Add "jax.distributed.initialize()" test

#1025 guptaaka closed 1 week ago
0
Support safetensors checkpoints [Do not merge - WIP]

#1024 richjames0 closed 1 week ago
1
Update checkpoint order for offload purpose

#1023 RissyRan closed 1 week ago
0
Gemma2 fix passing attn_logits_soft_cap config

#1022 gagika closed 1 week ago
0
Why logit checker has such a high tolerance?

#1021 hugoabonizio opened 1 week ago
0
Update maxtext_xpk_runner.py to show how to use vertex ai integrations

#1020 Obliviour opened 1 week ago
0
Support mixed-precision quantization configuration on AqtEinsum

#1019 lenscloth closed 1 week ago
1
[WIP] Support DiLoCo training

#1018 jonb377 opened 2 weeks ago
0
optimizations for offline mlperf inference

#1017 sixiang-google closed 1 week ago
2
Expose all SplashAttention tunable parameters in the workload.

#1016 vanbasten23 closed 1 week ago
2
Change sharding annotation for activation_embed_and_logits_batch

#1015 khatwanimohit closed 2 weeks ago
0
[DO NOT SUBMIT] Rotate via shmap + ppermute to test new mask idea

#1014 gobbleturk opened 2 weeks ago
0
fix data input on single CPU host

#1013 aireenmei closed 2 weeks ago
1
Correct vocab size for 8x22b

#1012 RissyRan closed 2 weeks ago
0
fix grain ckpt with new orbax

#1011 aireenmei closed 2 weeks ago
0
Name the initialize state module

#1010 gobbleturk closed 2 weeks ago
0
[WIP] Allow PP + megablox

#1009 gobbleturk opened 2 weeks ago
0