 AlpinDale
|
22305c91e9
refactor _prepare_model_input_tensor and attn metadata builder for most backends
|
hai 5 meses |
|
AlpinDale
|
9d7beaa5b9
chore: separate kv_scale into k_scale and v_scale
|
hai 5 meses |
|
AlpinDale
|
2105e4fd6b
feat: correctly invoke prefill & decode kernels for cross-attention
|
hai 5 meses |
 AlpinDale
|
405bb74612
Control plane comms refactor (#573)
|
hai 6 meses |
|
AlpinDale
|
ab7f4ed6e5
chore: revert commit for removing unnecessary copies in flash attn backend
|
hai 6 meses |
|
AlpinDale
|
156f577f79
feat: switch from `PYBIND11_MODULE` to `TORCH_LIBRARY` (#569)
|
hai 6 meses |
|
AlpinDale
|
75f97bc25d
bump flash-attn to remove unnecessary copies in the backend
|
hai 6 meses |
|
AlpinDale
|
696f2cd59c
add phi3_small support with blocksparse attention
|
hai 6 meses |
|
AlpinDale
|
b8b63eb5ca
fix head_size check for flash attention backend
|
hai 6 meses |
|
AlpinDale
|
93cffaf446
add flash_attn back
|
hai 6 meses |
|
AlpinDale
|
0c15965621
fix fp8 kv
|
hai 6 meses |
|
AlpinDale
|
a94de94c44
refactor: combine the prefill and decode into a single API (#553)
|
hai 6 meses |
|
AlpinDale
|
01190e5049
use flash attention for the decoding phase
|
hai 6 meses |
|
AlpinDale
|
50b7c13db0
refactor: attention selector (#552)
|
hai 6 meses |
|
AlpinDale
|
d11d68f4e6
switch to vllm-flash-attn
|
hai 6 meses |
|
AlpinDale
|
8b56dc4347
dict -> torch.Tensor for blocks_to_swap
|
hai 6 meses |
|
AlpinDale
|
3a0d1c7705
add get_name method to attention backends
|
hai 6 meses |
|
AlpinDale
|
21ce19b3ea
blocks_to_copy dict -> torch.Tensor
|
hai 6 meses |
|
AlpinDale
|
35ae01d7ba
refactor: attention metadata term
|
hai 6 meses |
|
AlpinDale
|
aa15d3dc0f
sliding window in prefix prefill kernel
|
hai 6 meses |
|
AlpinDale
|
fca911ee0a
vLLM Upstream Sync (#526)
|
hai 7 meses |
|
AlpinDale
|
66b7bc4415
sliding window in prefix kernel
|
hai 8 meses |
|
AlpinDale
|
9d81716bfd
[v0.5.3] Release Candidate (#388)
|
hai 9 meses |
