Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# Per-Head Mixed-Precision KV Cache Compression
|
| 2 |
|
| 3 |
Calibrate once. Pack truly. Same quality.
|
|
@@ -199,4 +223,4 @@ Step 3 — Results
|
|
| 199 |
|
| 200 |
MIT. Free to use, modify, and distribute.
|
| 201 |
|
| 202 |
-
Built in one week on an A100 SXM4 40GB. Questions, issues, and PRs welcome.
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
datasets:
|
| 4 |
+
- Salesforce/wikitext
|
| 5 |
+
language:
|
| 6 |
+
- en
|
| 7 |
+
metrics:
|
| 8 |
+
- perplexity
|
| 9 |
+
base_model:
|
| 10 |
+
- mistralai/Mistral-7B-Instruct-v0.3
|
| 11 |
+
- meta-llama/Meta-Llama-3-8B-Instruct
|
| 12 |
+
tags:
|
| 13 |
+
- quantization
|
| 14 |
+
- kv-cache
|
| 15 |
+
- llm-inference
|
| 16 |
+
- cuda
|
| 17 |
+
- triton
|
| 18 |
+
- memory-efficient
|
| 19 |
+
- mitral
|
| 20 |
+
- llama
|
| 21 |
+
- inference-optimization
|
| 22 |
+
- 4-bit
|
| 23 |
+
- mixed-precision
|
| 24 |
+
---
|
| 25 |
# Per-Head Mixed-Precision KV Cache Compression
|
| 26 |
|
| 27 |
Calibrate once. Pack truly. Same quality.
|
|
|
|
| 223 |
|
| 224 |
MIT. Free to use, modify, and distribute.
|
| 225 |
|
| 226 |
+
Built in one week on an A100 SXM4 40GB. Questions, issues, and PRs welcome.
|