Skip to content

Conversation

gmlwns2000
Copy link
Collaborator

WIP: AMD Support and Watchdog

Comment on lines 8 to 79
"using_extend": false,
"dense_layers": [0, 1, 2, 47, 46, 45],
"mask_refresh_interval": [96],
"mask_refresh_interval": [96, 32, 16],
"layers": [
{
"sliding_window_size": 1024,
"sliding_window_size_for_masking_step": [1024, 1024, 1024],
"second_stage_k": 1024,
"second_stage_k": 2048,
"sink_token_size": 1024,
"sa_extend_backend": "self_extend",
"stages": [ { } ]
"stages": [
{
"stage_block_size_q":128,
"stage_block_stride_q":4,
"stage_chunk_size":256,
"stage_k":null,
"stage_stride":1,
"using_landmark":false
},
{
"stage_block_size_q":64,
"stage_block_stride_q":1,
"stage_chunk_size":32,
"stage_k":65536,
"stage_stride":1,
"using_landmark":false
},
{
"stage_block_size_q":64,
"stage_block_stride_q":1,
"stage_chunk_size":8,
"stage_k":8192,
"stage_stride":1,
"using_landmark":false
}
]
},
{
"sliding_window_size": 1024,
"sliding_window_size_for_masking_step": [1024, 1024, 1024],
"second_stage_k": 1024,
"second_stage_k": 2048,
"sink_token_size": 1024,
"sa_extend_backend": "self_extend",
"scan_extend_backend": "none",
"stages": [ { } ]
"stages": [
{
"stage_block_size_q":128,
"stage_block_stride_q":4,
"stage_chunk_size":256,
"stage_k":null,
"stage_stride":1,
"using_landmark":false
},
{
"stage_block_size_q":64,
"stage_block_stride_q":1,
"stage_chunk_size":32,
"stage_k":65536,
"stage_stride":1,
"using_landmark":false
},
{
"stage_block_size_q":64,
"stage_block_stride_q":1,
"stage_chunk_size":8,
"stage_k":8192,
"stage_stride":1,
"using_landmark":false
}
]
}
],
Copy link
Contributor

@kbumsik kbumsik Oct 16, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I suppose it is used for qwen3.

So is this PR also fixes our qwen3 + nvidia? To me this PR can't be ignored even if we don't need AMD.

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.

2 participants