add bei specfic user migrations in trt-llm config (#1361)

* add bei specfic user migrations

* ruff fmt

truss/base/trt_llm_config.py

@@ -11,6 +11,8 @@
 from huggingface_hub.utils import validate_repo_id
 from pydantic import BaseModel, PydanticDeprecatedSince20, model_validator, validator
 
+from truss.base.constants import BEI_REQUIRED_MAX_NUM_TOKENS
+
 logger = logging.getLogger(__name__)
 # Suppress Pydantic V1 warnings, because we have to use it for backwards compat.
 warnings.filterwarnings("ignore", category=PydanticDeprecatedSince20)
@@ -119,6 +121,12 @@ class TrussTRTLLMBuildConfiguration(BaseModel):
     class Config:
         extra = "forbid"
 
+    def __init__(self, **data):
+        super().__init__(**data)
+        self._validate_kv_cache_flags()
+        self._validate_speculator_config()
+        self._bei_specfic_migration()
+
     @validator("max_beam_width")
     def check_max_beam_width(cls, v: int):
         if isinstance(v, int):
@@ -128,6 +136,29 @@ def check_max_beam_width(cls, v: int):
                 )
         return v
 
+    def _bei_specfic_migration(self):
+        """performs embedding specfic optimizations (no kv-cache, high batch size)"""
+        if self.base_model == TrussTRTLLMModel.ENCODER:
+            # Encoder specific settings
+            logger.info(
+                f"Your setting of `build.max_seq_len={self.max_seq_len}` is not used and "
+                "automatically inferred from the model repo config.json -> `max_position_embeddings`"
+            )
+
+            if self.max_num_tokens < BEI_REQUIRED_MAX_NUM_TOKENS:
+                logger.warning(
+                    f"build.max_num_tokens={self.max_num_tokens}, upgrading to {BEI_REQUIRED_MAX_NUM_TOKENS}"
+                )
+                self.max_num_tokens = BEI_REQUIRED_MAX_NUM_TOKENS
+            self.plugin_configuration.paged_kv_cache = False
+            self.plugin_configuration.use_paged_context_fmha = False
+
+            if "_kv" in self.quantization_type.value:
+                raise ValueError(
+                    "encoder does not have a kv-cache, therefore a kv specfic datatype is not valid"
+                    f"you selected build.quantization_type {self.quantization_type}"
+                )
+
     def _validate_kv_cache_flags(self):
         if not self.plugin_configuration.paged_kv_cache and (
             self.plugin_configuration.use_paged_context_fmha
@@ -176,11 +207,6 @@ def max_draft_len(self) -> Optional[int]:
             return self.speculator.num_draft_tokens
         return None
 
-    def __init__(self, **data):
-        super().__init__(**data)
-        self._validate_kv_cache_flags()
-        self._validate_speculator_config()
-
 
 class TrussSpeculatorConfiguration(BaseModel):
     speculative_decoding_mode: TrussSpecDecMode = TrussSpecDecMode.DRAFT_EXTERNAL