Spaces:
Running
Running
Expose conditional KV calculator in Space metadata
Browse files
README.md
CHANGED
|
@@ -19,6 +19,7 @@ tags:
|
|
| 19 |
- visualization
|
| 20 |
- mixture-of-experts
|
| 21 |
- attention
|
|
|
|
| 22 |
- recurrent-memory
|
| 23 |
- qwen
|
| 24 |
- flint
|
|
@@ -43,6 +44,8 @@ Explore two different architectures with real Q8_0 tensor captures: Flint's spar
|
|
| 43 |
|
| 44 |
The page also calculates raw f16 full-attention K/V storage across context lengths if a runtime enables caching. Nanbeige's released config sets `use_cache: false`, so this is a conditional calculation rather than measured default memory use.
|
| 45 |
|
|
|
|
|
|
|
| 46 |
[Read the architecture explainer](https://huggingface.co/blog/alphabell/nanbeige-qwen-layer-map) for pinned sources and a matched runtime test that could measure how context length affects each model.
|
| 47 |
|
| 48 |
## Five measured experiments
|
|
|
|
| 19 |
- visualization
|
| 20 |
- mixture-of-experts
|
| 21 |
- attention
|
| 22 |
+
- kv-cache
|
| 23 |
- recurrent-memory
|
| 24 |
- qwen
|
| 25 |
- flint
|
|
|
|
| 44 |
|
| 45 |
The page also calculates raw f16 full-attention K/V storage across context lengths if a runtime enables caching. Nanbeige's released config sets `use_cache: false`, so this is a conditional calculation rather than measured default memory use.
|
| 46 |
|
| 47 |
+
[Try the conditional 32K calculation](https://alphabell-model-observatory.static.hf.space/architecture/index.html?tokens=32768#kv-title) and move the slider to share another context length.
|
| 48 |
+
|
| 49 |
[Read the architecture explainer](https://huggingface.co/blog/alphabell/nanbeige-qwen-layer-map) for pinned sources and a matched runtime test that could measure how context length affects each model.
|
| 50 |
|
| 51 |
## Five measured experiments
|