diff --git a/chart/model-values.yaml b/chart/model-values.yaml
index b331dcf..2ba0a4b 100644
--- a/chart/model-values.yaml
+++ b/chart/model-values.yaml
@@ -365,16 +365,29 @@ deployments:
       ESRGAN_MODEL_IDS: |-
         RealESRGAN_x2plus
 
-  - name: "common-llms-sealion"
-    image: "crgooeyprodwestus1.azurecr.io/gooey-gpu-common:5"
+  - name: "common-llms-sealion-v2"
+    image: "crgooeyprodwestus1.azurecr.io/gooey-gpu-common:7"
+    limits_gpu: "30Gi"
     limits:
-      memory: "45Gi"
+      memory: "80Gi"  # (220 / 80) * 30
       cpu: "2"
     env:
       IMPORTS: |-
         common.llms
       LLM_MODEL_IDS: |-
-        aisingapore/sea-lion-7b-instruct
+        aisingapore/llama3-8b-cpt-sea-lionv2-instruct
+
+  - name: "common-llms-sarvam-2b"
+    image: "crgooeyprodwestus1.azurecr.io/gooey-gpu-common:7"
+    limits_gpu: "6Gi"
+    limits:
+      memory: "16Gi"  # (220 / 80) * 6
+      cpu: "2"
+    env:
+      IMPORTS: |-
+        common.llms
+      LLM_MODEL_IDS: |-
+        sarvamai/sarvam-2b-v0.5
 
 ## Dependencies
 nfs-server-provisioner:
diff --git a/common/llms.py b/common/llms.py
index 133eec7..49ec697 100644
--- a/common/llms.py
+++ b/common/llms.py
@@ -6,6 +6,7 @@
 import transformers
 from pydantic import BaseModel
 from transformers import AutoTokenizer
+from transformers.models.auto.tokenization_auto import get_tokenizer_config
 
 import gooey_gpu
 from celeryconfig import app, setup_queues
@@ -14,10 +15,11 @@
 class PipelineInfo(BaseModel):
     model_id: str
     seed: int = None
+    fallback_chat_template_from: str | None
 
 
 class LLMChatInputs(BaseModel):
-    messages: typing.List[dict]
+    text_inputs: typing.List[dict] | str
     max_new_tokens: int
     stop_strings: typing.Optional[typing.List[str]]
     temperature: float = 1
@@ -33,8 +35,15 @@ class LLMChatOutput(BaseModel):
 @gooey_gpu.endpoint
 def llm_chat(pipeline: PipelineInfo, inputs: LLMChatInputs) -> LLMChatOutput:
     pipe = load_pipe(pipeline.model_id)
-    return pipe(
-        inputs.messages,
+
+    if pipeline.fallback_chat_template_from and not pipe.tokenizer.chat_template:
+        # if the tokenizer does not have a chat template, use the provided fallback
+        config = get_tokenizer_config(pipeline.fallback_chat_template_from)
+        pipe.tokenizer.chat_template = config.get("chat_template")
+
+    # for a list of parameters, see https://huggingface.co/docs/transformers/en/main_classes/text_generation#transformers.GenerationConfig
+    ret = pipe(
+        inputs.text_inputs,
         max_new_tokens=inputs.max_new_tokens,
         stop_strings=inputs.stop_strings,
         temperature=inputs.temperature,
@@ -44,17 +53,26 @@ def llm_chat(pipeline: PipelineInfo, inputs: LLMChatInputs) -> LLMChatOutput:
         eos_token_id=pipe.tokenizer.eos_token_id,
     )[0]
 
+    # strip stop strings & eos token from final output
+    for s in (inputs.stop_strings or []) + [pipe.tokenizer.eos_token]:
+        ret["generated_text"] = ret["generated_text"].split(s, 1)[0]
+
+    return ret
+
 
 @lru_cache
-def load_pipe(model_id: str):
+def load_pipe(model_id: str) -> transformers.TextGenerationPipeline:
     print(f"Loading llm model {model_id!r}...")
     # this should return a TextGenerationPipeline
-    pipe = transformers.pipeline(
-        "text-generation",
-        model=model_id,
-        device=gooey_gpu.DEVICE_ID,
-        torch_dtype=torch.float16,
-        trust_remote_code=True,
+    pipe = typing.cast(
+        transformers.TextGenerationPipeline,
+        transformers.pipeline(
+            "text-generation",
+            model=model_id,
+            device=gooey_gpu.DEVICE_ID,
+            torch_dtype=torch.float16,
+            trust_remote_code=True,
+        ),
     )
     if not pipe.tokenizer:
         pipe.tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
diff --git a/scripts/run-dev.sh b/scripts/run-dev.sh
index f3bd08a..4350b50 100755
--- a/scripts/run-dev.sh
+++ b/scripts/run-dev.sh
@@ -74,7 +74,7 @@ docker run \
     RealESRGAN_x2plus
   "\
   -e LLM_MODEL_IDS="
-    aisingapore/sea-lion-7b-instruct
+    aisingapore/llama3-8b-cpt-sea-lionv2-instruct
   "\
   -e C_FORCE_ROOT=1 \
   -e BROKER_URL=${BROKER_URL:-"amqp://"} \