Skip to content

Commit 3991768

Browse files
committed
Fix diagram: replace Unicode box chars with ASCII (GitHub rendering)
1 parent fa156fc commit 3991768

2 files changed

Lines changed: 18 additions & 18 deletions

File tree

README.ko.md

Lines changed: 9 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -27,15 +27,15 @@
2727
LLM 메모리의 병목은 모델 가중치가 아니라 **KV 캐시**입니다. 32K 컨텍스트에서 8B 모델의 KV 캐시는 **4GB** — 모델 자체보다 큽니다. 기존 엔진은 모두 KV를 FP16으로 저장합니다. 우리는 이것을 압축합니다.
2828

2929
```
30-
┌──────────────────────────────────────────────┐
31-
16GB Mac 메모리
32-
├────────────┬─────────────────────────────────┤
33-
│ 모델(4GB) │ KV (FP16)
34-
│ ██████████████ 8K ctx ← OOM
35-
├────────────┼─────────────────────────────────┤
36-
│ 모델(4GB) KV (4-bit) ██ →→ 350K ctx
37-
↑ 6.9배 작음
38-
└────────────┴─────────────────────────────────┘
30+
+------------+-------------------------------+
31+
| | 16GB Mac Memory |
32+
| Model(4GB) +-------------------------------+
33+
| | KV Cache (FP16) |
34+
| | [##############] 8K <-- OOM |
35+
+------------+-------------------------------+
36+
| Model(4GB) | KV (4-bit) [##] --> 350K ctx |
37+
| | 6.9x smaller |
38+
+------------+-------------------------------+
3939
```
4040

4141
## 결과

README.md

Lines changed: 9 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -27,15 +27,15 @@
2727
LLM memory is dominated by the **KV cache**, not model weights. At 32K context, a 8B model's KV cache consumes **4GB** — more than the model itself. Every existing engine stores KV in FP16. We compress it.
2828

2929
```
30-
┌──────────────────────────────────────────────┐
31-
16GB Mac Memory
32-
├────────────┬─────────────────────────────────┤
33-
│ Model(4GB) │ KV (FP16)
34-
│ ██████████████ 8K ctx ← OOM
35-
├────────────┼─────────────────────────────────┤
36-
Model(4GB) KV (4-bit) ██ →→ 350K ctx
37-
6.9x smaller
38-
└────────────┴─────────────────────────────────┘
30+
+------------+-------------------------------+
31+
| | 16GB Mac Memory |
32+
| Model(4GB) +-------------------------------+
33+
| | KV Cache (FP16) |
34+
| | [##############] 8K <-- OOM |
35+
+------------+-------------------------------+
36+
| Model(4GB) | KV (4-bit) [##] --> 350K ctx |
37+
| | 6.9x smaller |
38+
+------------+-------------------------------+
3939
```
4040

4141
## The Result

0 commit comments

Comments
 (0)