electron-rare/prima-cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

doc: update doc for q2_k

Browse Source
This commit is contained in:
DandinPower
2025-11-03 19:44:38 +08:00
parent 9e1a4f2942
commit f90029e3c3
1 changed files with 2 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CHANGES.md
+2 -1
View File
@@ -258,6 +258,7 @@ A new CLI argument (`--comm_datatype TYPE`) sets the communication data type. Su
- f32 (default, no compression)
- q8_0 (8-bit block quantization)
- q4_0 (4-bit block quantization)
- q2_k (2-bit k quantization)
- f32_sparsity (no quantization, but allows `--comm_sparse_percentage` to select the top-k features for each token based on the specified sparsity percentage)
### Communication Logic
@@ -265,7 +266,7 @@ A new CLI argument (`--comm_datatype TYPE`) sets the communication data type. Su
- **Send path**
- If `f32`, tensors are transmitted directly.
- If `q8_0` or `q4_0`, tensors are quantized before transmission.
- If `q8_0`, `q4_0` or `q2_k`, tensors are quantized before transmission.
- If `f32_sparsity`, tensors are pruned before transmission.
- **Receive path**