Skip to content

Update spiceai from EricLBuehler/candle#6

Merged
Jeadie merged 12 commits into
spiceaifrom
jeadie/25-04-15/upstream-spiceai
Apr 15, 2025
Merged

Update spiceai from EricLBuehler/candle#6
Jeadie merged 12 commits into
spiceaifrom
jeadie/25-04-15/upstream-spiceai

Conversation

@Jeadie
Copy link
Copy Markdown

@Jeadie Jeadie commented Apr 15, 2025

Outstanding diff to ericLbuehler is EricLBuehler@fd28f08

EricLBuehler and others added 10 commits March 11, 2025 22:07
* Add FlashMLA

* Add flash attn rust ffi

* Automatic computation of mla metadata

* Add some shape checks

* Fix .a name

* Fix linking name mha_fwd_kvcache_mla

* extern "C"

* Handle CUDA_NVCC_FLAGS

* Include flash_fwd_mla_kernel.h again

* Include flash_fwd_mla_kernel.h only once

* Tweak

* Add flash_fwd_mla_kernel.h

* Use cute::bfloat16_t

* Move CUDA_NVCC_FLAGS to last

* Fix reshape

* Only k_c_k_pe cache, no k/v cache

* Fix passing head_size_v

* Remove check for "v"

* Fix out shape

* out-accum should be f32

* Remove references to flashattnv3

* Add test

* Some fixes

* Use repeat interleave

* Maybe some progress...

* Tests pass!

* Move to excluded
* Support sdpa with mask, causal

* Properly handle softcapping
@Jeadie Jeadie self-assigned this Apr 15, 2025
@Jeadie
Copy link
Copy Markdown
Author

Jeadie commented Apr 15, 2025

Failing in EricLBuehler: EricLBuehler@bca0107

@Jeadie Jeadie merged commit 1b02ddb into spiceai Apr 15, 2025
1 check passed
@Jeadie Jeadie mentioned this pull request Apr 15, 2025
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.

4 participants