Upload llmdolphin.py

llmdolphin.py

@@ -27,16 +27,25 @@ llm_models = {
     "Nemo-12B-Marlin-v5-Q4_K_M.gguf": ["starble-dev/Nemo-12B-Marlin-v5-GGUF", MessagesFormatterType.CHATML],
     "Nemo-12B-Marlin-v7.Q4_K_M.gguf": ["mradermacher/Nemo-12B-Marlin-v7-GGUF", MessagesFormatterType.MISTRAL],
     "NemoDori-v0.2-Upscaled.1-14B.Q4_K_M.gguf": ["mradermacher/NemoDori-v0.2-Upscaled.1-14B-GGUF", MessagesFormatterType.MISTRAL],
+    "Fireball-12B-v1.0.i1-Q4_K_M.gguf": ["mradermacher/Fireball-12B-v1.0-i1-GGUF", MessagesFormatterType.MISTRAL],
     "Fireball-Mistral-Nemo-Base-2407-sft-v2.2a.Q4_K_M.gguf": ["mradermacher/Fireball-Mistral-Nemo-Base-2407-sft-v2.2a-GGUF", MessagesFormatterType.MISTRAL],
+    "StorieCreative.i1-Q4_K_S.gguf": ["mradermacher/StorieCreative-i1-GGUF", MessagesFormatterType.MISTRAL],
     "Deutscher-Pantheon-12B.Q4_K_M.gguf": ["mradermacher/Deutscher-Pantheon-12B-GGUF", MessagesFormatterType.MISTRAL],
     "Trinas_Nectar-8B-model_stock.i1-Q4_K_M.gguf": ["mradermacher/Trinas_Nectar-8B-model_stock-i1-GGUF", MessagesFormatterType.MISTRAL],
+    "Loki-v5.2.Q5_K_M.gguf": ["mradermacher/Loki-v5.2-GGUF", MessagesFormatterType.MISTRAL],
+    "Loki-v5.1.Q5_K_M.gguf": ["mradermacher/Loki-v5.1-GGUF", MessagesFormatterType.MISTRAL],
+    "GracieRP-freefallenLora-Gemma2-Inst-9B.i1-Q4_K_M.gguf": ["mradermacher/GracieRP-freefallenLora-Gemma2-Inst-9B-i1-GGUF", MessagesFormatterType.ALPACA],
+    "mistral-nemo-gutenberg-12B-v4.Q4_K_M.gguf": ["mradermacher/mistral-nemo-gutenberg-12B-v4-GGUF", MessagesFormatterType.MISTRAL],
+    "FunkyMerge-12b-0.1.Q4_K_M.gguf": ["mradermacher/FunkyMerge-12b-0.1-GGUF", MessagesFormatterType.MISTRAL],
     "NemoMix-Unleashed-12B-Q4_K_M.gguf": ["bartowski/NemoMix-Unleashed-12B-GGUF", MessagesFormatterType.MISTRAL],
     "IceTea21EnergyDrinkRPV13.Q4_K_S.gguf": ["mradermacher/IceTea21EnergyDrinkRPV13-GGUF", MessagesFormatterType.MISTRAL],
     "MegaBeam-Mistral-7B-512k-Q5_K_M.gguf": ["bartowski/MegaBeam-Mistral-7B-512k-GGUF", MessagesFormatterType.MISTRAL],
+    "azur-8b-model_stock-q4_k_m.gguf": ["DreadPoor/Azur-8B-model_stock-Q4_K_M-GGUF", MessagesFormatterType.LLAMA_3],
     "Chronos-Gold-12B-1.0-Q4_K_M.gguf": ["bartowski/Chronos-Gold-12B-1.0-GGUF", MessagesFormatterType.MISTRAL],
     "L3.1-Romes-Ninomos-Maxxing.Q5_K_M.gguf": ["mradermacher/L3.1-Romes-Ninomos-Maxxing-GGUF", MessagesFormatterType.LLAMA_3],
     "mistral-nemo-minitron-8b-base-q4_k_m.gguf": ["Daemontatox/Mistral-NeMo-Minitron-8B-Base-Q4_K_M-GGUF", MessagesFormatterType.MISTRAL],
     "Nokstella_coder-8B-model_stock.i1-Q4_K_S.gguf": ["mradermacher/Nokstella_coder-8B-model_stock-i1-GGUF", MessagesFormatterType.LLAMA_3],
+    "vtion_model_v1.Q5_K_M.gguf": ["mradermacher/vtion_model_v1-GGUF", MessagesFormatterType.LLAMA_3],
     "storiecreative-q5_k_m.gguf": ["ClaudioItaly/StorieCreative-Q5_K_M-GGUF", MessagesFormatterType.MISTRAL],
     "L3.1-gramamax.Q5_K_M.gguf": ["mradermacher/L3.1-gramamax-GGUF", MessagesFormatterType.MISTRAL],
     "Evolutionstory128.Q5_K_M.gguf": ["mradermacher/Evolutionstory128-GGUF", MessagesFormatterType.CHATML],
@@ -668,7 +677,7 @@ def get_raw_prompt(msg: str):
     return re.sub(r'[*/:_"#\n]', ' ', ", ".join(m)).lower() if m else ""
 
 
-@spaces.GPU(duration=30)
+@spaces.GPU(duration=60)
 def dolphin_respond(
     message: str,
     history: list[tuple[str, str]],
@@ -761,7 +770,7 @@ def dolphin_parse(
     return ", ".join(prompts), gr.update(interactive=True), gr.update(interactive=True)
 
 
-@spaces.GPU(duration=30)
+@spaces.GPU(duration=60)
 def dolphin_respond_auto(
     message: str,
     history: list[tuple[str, str]],
@@ -788,7 +797,7 @@ def dolphin_respond_auto(
         flash_attn=True,
         n_gpu_layers=81, # 81
         n_batch=1024,
-        n_ctx=4096, #8192
+        n_ctx=8192, #8192
     )
     provider = LlamaCppPythonProvider(llm)