Update GPU performance tuning and documentation for context length

- Enhanced the `OllamaTune` class to include a new `context_length` attribute, improving the configuration for different GPU tiers.
- Updated performance tuning logic to set appropriate context lengths for low, mid, high, and ultra tiers, ensuring optimal resource allocation.
- Modified installation scripts to reflect the new context length settings, enhancing the installation process for Ollama.
- Revised documentation to include context length details in the GPU performance table, providing clearer guidance for users.
- Added tests to validate the correct context length settings in various scenarios, ensuring robustness in performance tuning.

src/gpu_rent/perf_tiers.py

@@ -30,6 +30,7 @@ class OllamaTune:
     max_loaded_models: int
     kv_cache_type: str | None
     gpu_overhead_bytes: int
+    context_length: int
     notes: str
 
 
@@ -72,23 +73,27 @@ def ollama_tune_for(info: GpuInfo) -> OllamaTune:
         overhead = 20 * 1024**3
         keep = "30m"
         kv = "q8_0"
-        note = "ultra: flash+q8 KV, 20GiB reserved for Swarm, keep 30m"
+        ctx = 32768
+        note = "ultra: flash+q8 KV, 20GiB reserved for Swarm, ctx 32k, keep 30m"
     elif info.tier == TIER_HIGH:
         overhead = 14 * 1024**3
         keep = "15m"
         kv = "q8_0"
-        note = "high: flash+q8 KV, 14GiB reserved for Swarm, keep 15m"
+        ctx = 16384
+        note = "high: flash+q8 KV, 14GiB reserved for Swarm, ctx 16k, keep 15m"
     elif info.tier == TIER_MID:
         overhead = 10 * 1024**3
         keep = "5m"
         kv = "q8_0"
-        note = "mid: flash+q8 KV, 10GiB reserved for Swarm, keep 5m"
+        ctx = 16384
+        note = "mid: flash+q8 KV, 10GiB reserved for Swarm, ctx 16k, keep 5m"
     else:
         overhead = 6 * 1024**3
         keep = "2m"
         kv = "q4_0"
+        ctx = 8192
         flash = False  # prefer stability on tiny cards
-        note = "low: conservative, 6GiB reserved, short keep-alive"
+        note = "low: conservative, 6GiB reserved, ctx 8k, short keep-alive"
 
     return OllamaTune(
         flash_attention=flash,
@@ -97,6 +102,7 @@ def ollama_tune_for(info: GpuInfo) -> OllamaTune:
         max_loaded_models=1,
         kv_cache_type=kv,
         gpu_overhead_bytes=overhead,
+        context_length=ctx,
         notes=note,
     )
 
@@ -139,6 +145,7 @@ def ollama_env_lines(tune: OllamaTune) -> list[str]:
         f"Environment=OLLAMA_MAX_LOADED_MODELS={tune.max_loaded_models}",
         f"Environment=OLLAMA_KEEP_ALIVE={tune.keep_alive}",
         f"Environment=OLLAMA_GPU_OVERHEAD={tune.gpu_overhead_bytes}",
+        f"Environment=OLLAMA_CONTEXT_LENGTH={tune.context_length}",
     ]
     if tune.flash_attention:
         lines.append("Environment=OLLAMA_FLASH_ATTENTION=1")