Mathis Felardos

Replace pacmd calls with pactl to support Pipewire [final version]

1

Hello, I continued the work from @EgZvor (thank you!) in #818 and patched the remaining issue with default-sink in pactl. I used this version for the last two weeks so...

fastapi-jwt should have other jwt backends besides python-jose. Authlib seems a good candidate.

1

Hello @k4black, First, I would like to thank you for your project. I sincerely believe that this project should be merged into the main branch of fastapi. Now, I think...

Feature request: Dark theme (with a suggestion of what the CSS code could be)

10

Hi, It would be awesome to add a Dark Theme (like Tree Style Tab). It could be a simple option. So far I am overriding the css in the `chrome/userContent.css`...

[Bugfix][Disaggregated] Add a check in send_kv_caches_and_hidden_states and fix the reshape of the KVCache

1

Hello, This PR adds this in `simple_connector.py::send_kv_caches_and_hidden_states`: ``` head_size = getattr(model_config, "head_dim", int(hidden_size // num_attention_heads)) ``` Indeed, the assumption that `head_size == hidden_size // num_attention_heads` is not always true. I...

[Config][Disaggregated] Add timeout configuration for the torch.store and add KVTransferConfig.kv_connector_extra_config

1

Hello, Adding a kv_connector_extra_config will help custom connectors to use it to hack around if they need. I'm using it within the SimpleBuffer/PyNcclPipe to set the `torch.distributed` Store's timeout (in...

[bugfix] avoid NIXL_ERR_REMOTE_DISCONNECT in nixl_connector when Prefill dies

2

## Purpose Avoid NIXL_ERR_REMOTE_DISCONNECT in `NixlConnectorWorker._pop_done_transfers(...)` for a Decode instance when a Prefill dies. This is the stacktrace we experienced: ``` File "/opt/venv/lib/python3.12/site-packages/vllm/v1/worker/gpu_model_runner.py", line 2517, in execute_model self.maybe_get_kv_connector_output(scheduler_output) as ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^...

DeepSeek KVCache computation has a regression in the kv_cache_calculator

1

DeepSeek KVCache computation has a regression in the kv_cache_calculator: The [original PR ](https://github.com/LMCache/LMCache/pull/671) used to show the proper value: cc @zzhbrr, @zhuohangu