🐛 fix lm_head weight mapping (#76)

#### Motivation

Fixing an internal issue where
`ml6team/keyphrase-generation-t5-small-inspec` and the `t5-small` models
were failing due to `RuntimeError: weight lm_head.weight does not exist`
error.

#### Modifications

We added the `lm_head.weight` alias to the list of known aliases for
`t5` type models. (It was discovered within the `metadata` of the tensor
file during inspection)

```
"lm_head.weight":["decoder.embed_tokens.weight"]
```

#### Result

Now we're able to launch `t5` based
`ml6team/keyphrase-generation-t5-small-inspec` and `t5-small` models
using `tgis_native`!
```
DEPLOYMENT_FRAMEWORK=tgis_native MODEL_NAME=ml6team/keyphrase-generation-t5-small-inspec text-generation-launcher
```

---------

Signed-off-by: Prashant Gupta <[email protected]>
Signed-off-by: Joe Runde <[email protected]>
Co-authored-by: Joe Runde <[email protected]>

Author: prashantgupta24

server/text_generation_server/inference_engine/tgis_native.py

@@ -64,7 +64,10 @@ def __init__(
             model_class = BloomForCausalLM
 
         elif model_type == "t5":
-            aliases = {"shared.weight": ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight"]}
+            aliases = {
+                "shared.weight": ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight"], 
+                "lm_head.weight": ["decoder.embed_tokens.weight"],
+            }
             from text_generation_server.models.custom_modeling.t5_modeling import T5ForConditionalGeneration
             model_class = T5ForConditionalGeneration