[Paged KV][Doc] Clarify memory env var for two path in readme (#140)

Signed-off-by: ran <hzz5361@psu.edu>
Signed-off-by: Yuan Lik Xun <lxyuan0420@gmail.com>
Co-authored-by: Yuan Lik Xun <lxyuan0420@gmail.com>

Author: WindChimeRan

README.md

@@ -101,3 +101,17 @@ Environment variables for customization:
 | `VLLM_METAL_MODELSCOPE_CACHE` | None | Specify the absolute path of the local model |
 | `VLLM_METAL_PREFIX_CACHE` | (unset) | Set to enable prefix caching for shared prompt reuse |
 | `VLLM_METAL_PREFIX_CACHE_FRACTION` | `0.05` | Fraction of MLX working set for prefix cache (0, 1] |
+
+
+## Paged KV vs MLX KV memory settings
+
+- MLX path (`VLLM_METAL_USE_PAGED_ATTENTION=0`): `VLLM_METAL_MEMORY_FRACTION` must be `auto`.
+- Paged KV path (`VLLM_METAL_USE_PAGED_ATTENTION=1`): `VLLM_METAL_MEMORY_FRACTION` can be `auto` or a numeric fraction in `(0, 1]`.
+- For paged KV with `VLLM_METAL_MEMORY_FRACTION=auto`, vllm-metal uses a default fraction of `0.9`.
+
+`VLLM_METAL_MEMORY_FRACTION` | `VLLM_METAL_USE_PAGED_ATTENTION` | Valid? | Notes
+-- | -- | -- | --
+`auto` | `0` | Yes | MLX path (default)
+`auto` | `1` | Yes | Paged KV path; defaults to 0.9 internally
+`0.7` | `1` | Yes | Paged KV path with explicit memory budget
+`0.7` | `0` | No | Explicit fraction without paged KV is invalid