nvMelissa

Results 29 comments of nvMelissa

RuntimeError and Socket Connection Failure when Benchmarking Gemma-7b with Micro Batch Size 1

Per triage meeting on 8/26, moved to @t-vi and assigned priority as P2.

sdpa_ex - Incorrect device in trace vs from actual computation

Per triage meeting, assigned to @IvanYashchuk to find right owner

Guard nvFuser-based tests with minimum required compute capability or BF16 capability

From Kevin: The problem is not in nvFuser but in trying to run BF16 in Arch prior to SM8.0 ( Ampere)

Incompatibility with HF Model Qwen2-1.5B - Tensor Indexing Error (1-D vs 2-D)

Triage review 9/16/24: Problem acknowledged. We welcome the community to update the meta function to solve the problem.

Add nvFuser support for thunder.torch.randn

From Kevin : There are no plans to support this at the moment. We can revisit if there is a model that has this requirement.

Thunder seems to use way more memory when `litgpt.Config.parallel_residual=True`

Add to Q4 planning

[PyTorch] CPU Overhead Micro-optimizations

@zhongbozhu - this is out of date with base branch . Woudl you please retry with latest ?

[JAX] cuBlasMp integration for CollectiveGemm custom op

Hi @denera - This needs to go into TE2.10. I see @phu0ngng 's last comment to add some tests. Please let us know if that was already done and if...

Feature fast cast-only mxfp8

Hello @Jianbing-D, thanks for your contribution! Would you please check why the PR is failing CI checks? If you don't find the reason, please request support from the reviewers. Thank...

thd format is not supported with hierarchical CP implementation yet

Hi @xrennvidia - thank you for the prompt response to @stormchasingg . Would you please provide an estimate when THD format is expected to be supported on so we can...