Add Kimi-k2-thinking and k2.5 and k2.6 checkpoint conversion support.#3768
Add Kimi-k2-thinking and k2.5 and k2.6 checkpoint conversion support.#3768copybara-service[bot] merged 1 commit intomainfrom
Conversation
|
🤖 Hi @gagika, I've received your request, and I'm working on it now! You can track my progress in the logs for more details. |
|
🤖 I'm sorry @gagika, but I was unable to process your request. Please see the logs for more details. |
|
🤖 Hi @gagika, I've received your request, and I'm working on it now! You can track my progress in the logs for more details. |
|
🤖 Hi @gagika, I've received your request, and I'm working on it now! You can track my progress in the logs for more details. |
|
🤖 I'm sorry @gagika, but I was unable to process your request. Please see the logs for more details. |
|
🤖 Hi @RissyRan, I've received your request, and I'm working on it now! You can track my progress in the logs for more details. |
|
🤖 I'm sorry @RissyRan, but I was unable to process your request. Please see the logs for more details. |
richjames0
left a comment
There was a problem hiding this comment.
Great testing Gagik. One nth comment
|
🤖 Hi @RissyRan, I've received your request, and I'm working on it now! You can track my progress in the logs for more details. |
|
🤖 Hi @RissyRan, I've received your request, and I'm working on it now! You can track my progress in the logs for more details. |
|
🤖 I'm sorry @RissyRan, but I was unable to process your request. Please see the logs for more details. |
shuningjin
left a comment
There was a problem hiding this comment.
Thank you for expanding support to multiple variants (K2-Thinking, K2.5, K2.6) on top of Kimi-K2! The int4 dequantization with unit tests, checkpoint conversion, and user guide updates all look great.
Description
kimi-k2.6-text) to the MaxText DeepSeek family layout.
compressed-tensors pack-quantized weights with per-group symmetric scales (group_size=32).
Tests
KL divergence = [0.00043962 0.00289483]: https://paste.googleplex.com/6351128208998400Checklist
Before submitting this PR, please make sure (put X in square brackets):
gemini-reviewlabel.