feat: update ibm_fms engine to support variant config overrides (IBM#58)

#### Motivation

The ibm_fms engine is being too strict on the "variant" selection. The
FMS get_model call requires the variant, but also supports overriding
configuration parameters via kwargs.

#### Modifications

Allows for support for additional FMS Calico models that may have
differences from the default variant config, such as the size of the
vocab.

Signed-off-by: Travis Johnson <[email protected]>

Author: tjohnson31415

server/text_generation_server/inference_engine/ibm_fms.py

@@ -1,15 +1,19 @@
 import torch
 from typing import Any, Optional, Dict
+from loguru import logger
 
 from fms.models import get_model, list_variants, __models as _fms_models
 from fms.models.hf import to_hf_api
-from fms_extras.models.calico import CalicoConfig
+from fms.utils.activation import __ACT_2_CLS as FMS_ACT_2_CLASS
 
 from text_generation_server.inference_engine.engine import BaseInferenceEngine
 
+# Register FMS model classes with HF
 from fms_extras.models.hf import register_fms_models
+
 register_fms_models()
 
+
 class InferenceEngine(BaseInferenceEngine):
     def __init__(
         self,
@@ -22,60 +26,143 @@ def __init__(
     ) -> None:
         # model_config comes in a as a Dict to support the late registration
         if (model_type := model_config["model_type"]) != "gpt_megatron":
-            raise ValueError(f"Unknown model type {model_type} passed to ibm_fms engine.")
+            raise ValueError(
+                f"Unknown model type {model_type} passed to ibm_fms engine."
+            )
 
         super().__init__(model_path, model_config)
 
-        variant_match = self._find_variant(model_config, 'calico')
+        # only calico supported currently
+        fms_architecture_name = "calico"
+
+        # model_config can override what is set by the variant
+        fms_config_dict = self._convert_model_config(model_config)
+
+        variant_match = self._find_variant(fms_config_dict, fms_architecture_name)
         if variant_match is None:
-            raise ValueError(f"Unable to determine model variant from model: {model_path}")
+            raise ValueError(
+                f"Unable to determine model variant for model: {model_path}"
+            )
 
         # get_model does not have a dtype parameter, setting the default dtype
         # reduces memory required to load the model compared to converting afterwards
         orig_dtype = torch.get_default_dtype()
         try:
             torch.set_default_dtype(dtype)
-            calico_model = get_model("calico", variant_match, model_path, source="megatron", device_type=self.device.type)
-
-            # the CalicoConfig does not persist token ids to the
-            # HFAdaptedConfig, so pass them along explicitly
-            self.model = to_hf_api(
-                calico_model,
-                pad_token_id=model_config['pad_token_id'],
-                bos_token_id=model_config['bos_token_id'],
-                eos_token_id=model_config['eos_token_id'],
-            ).requires_grad_(False).eval()
+            fms_model = get_model(
+                fms_architecture_name,
+                variant_match,
+                model_path,
+                source="megatron",
+                device_type=self.device.type,
+                **fms_config_dict,
+            )
+            # get_model does not persist token ids to the HFAdaptedConfig, so
+            # pass them along explicitly
+            self.model = (
+                to_hf_api(
+                    fms_model,
+                    pad_token_id=model_config["pad_token_id"],
+                    bos_token_id=model_config["bos_token_id"],
+                    eos_token_id=model_config["eos_token_id"],
+                )
+                .requires_grad_(False)
+                .eval()
+            )
         finally:
             torch.set_default_dtype(orig_dtype)
 
         # update the config to config object instead of a dict
         self._config = self.model.config
 
     @classmethod
-    def _find_variant(cls, model_config_dict: Dict, fms_architecture) -> Optional[str]:
+    def _convert_model_config(cls, model_config_dict: Dict) -> Dict:
+        # mapping between CalicoConfig attributes and keys in the model_config dict
+        act_fn_attr = "activation_fn"
+        fms_config_attr_to_config_json_key = {
+            "src_vocab_size": "vocab_size",
+            "emb_dim": "n_embd",
+            "norm_eps": "layer_norm_epsilon",
+            "nheads": "n_head",
+            "kvheads": "num_key_value_heads",
+            "nlayers": "n_layer",
+            "pad_id": "pad_token_id",
+            # No entry in config.json for hidden_growth_factor
+            # No entry in config.json for multiple_of
+            act_fn_attr: "activation_function",
+            # ignore p_dropout for inference
+            "max_expected_seq_len": "n_positions",
+        }
+        fms_config_dict = {
+            attr: model_config_dict[key]
+            for attr, key in fms_config_attr_to_config_json_key.items()
+            if key in model_config_dict
+        }
+
+        # the activation function name may need to be converted
+        if act_fn := fms_config_dict.get(act_fn_attr):
+            fms_config_dict[act_fn_attr] = cls._convert_activation_function_name(act_fn)
+
+        return fms_config_dict
+
+    @classmethod
+    def _convert_activation_function_name(cls, act_name: str) -> str:
+        """Attempts to find an FMS compatible activation function name
+
+        gpt_megatron models may use different names for the activation function
+        compared to FMS, specifically around whether "GLU" is indicated
+        explicitly.
+
+        Refer to the fms.utils.activation module to see supported names
+        """
+        glu_activation_function_mapping = {
+            "geglu": "gelu",
+            "miglu": "mish",
+            "mishglu": "mish",
+            "reglu": "relu",
+            "swiglu": "swish",
+        }
+        if act_name.endswith("_glu"):
+            fms_act_name = act_name.rstrip("_glu")
+        elif new_name := glu_activation_function_mapping.get(act_name):
+            fms_act_name = new_name
+        else:
+            fms_act_name = act_name
+
+        # ensure the final act name is supported by FMS
+        if fms_act_name not in FMS_ACT_2_CLASS:
+            raise ValueError(f"Unsupported activation function: {act_name}.")
+
+        return fms_act_name
+
+    @classmethod
+    def _find_variant(
+        cls, fms_config_dict: Dict, fms_architecture: str
+    ) -> Optional[str]:
         # get a list of variant configs to compare against the model_config_dict
         variant_map = {
             # HACK: extract the variant config from the closure created for the factory functions...
             v: _fms_models[fms_architecture][v].__closure__[0].cell_contents
             for v in list_variants(fms_architecture)
         }
-        for v, v_config in variant_map.items():
-            if cls._is_variant_compatible(model_config_dict, v_config):
-                return v
-        return None
 
-    @classmethod
-    def _is_variant_compatible(cls, model_config_dict: Dict, config: CalicoConfig) -> bool:
-        dict_key_to_attr = {
-            'vocab_size': 'src_vocab_size',
-            'n_embd': 'emb_dim',
-            'n_head': 'nheads',
-            'num_key_value_heads': 'kvheads',
-            'n_layer': 'nlayers',
-            'n_positions': 'max_expected_seq_len',
-            'pad_token_id': 'pad_id',
-        }
-        for key, attr in dict_key_to_attr.items():
-            if model_config_dict[key] != getattr(config, attr, None):
-                return False
-        return True
+        # attributes of the CalicoConfig that must exist and match to find a
+        # compatible "variant"
+        variant_attrs_to_check = [
+            "emb_dim",
+            "nheads",
+            "kvheads",
+            "nlayers",
+        ]
+        if not all(fms_config_dict.get(attr, None) for attr in variant_attrs_to_check):
+            raise ValueError(
+                f"Unable to find compatible variant, the following configurations must exist {variant_attrs_to_check}"
+            )
+
+        for v_name, v_config in variant_map.items():
+            if all(
+                fms_config_dict.get(attr) == getattr(v_config, attr)
+                for attr in variant_attrs_to_check
+            ):
+                return v_name
+        return None