backend: Extend the models config YAML file

Extend the models config YAML file to include model-specific
information.

Add model-specific details, such as the number of parameters, memory
footprint, and tensor type (e.g., FP32, BF16, etc.), along with a set
of default values used by Lumigator when calling the models.

Not all parameters are applicable to each model. For example, Hugging
Face models have default values for parameters like `max_length`, while
API models include parameters like `temperature`. In general, parameters
are model-specific, so there is no universal set of parameter names.

The defaults are inferred either by checking the documentation for each
model or by examining the configuration file on the Hugging Face Hub.

Closes #381

Signed-off-by: Dimitris Poulopoulos <[email protected]>

lumigator/python/mzai/backend/backend/models_config.yaml

@@ -1,35 +1,96 @@
 - name: facebook/bart-large-cnn
   uri: hf://facebook/bart-large-cnn
   description: BART is a large-sized model fine-tuned on the CNN Daily Mail dataset.
+  info:
+    parameters_count: 406M
+    tensor_type: F32
+    model_size: 1.63GB
+  default_parameters:
+    max_length: 142
+    min_length: 56
+    length_penalty: 2.0
+    early_stopping: true
+    no_repeat_ngram_size: 3
+    num_beams: 4
 
 - name: mikeadimech/longformer-qmsum-meeting-summarization
   uri: hf://mikeadimech/longformer-qmsum-meeting-summarization
   description: Longformer is a transformer model that is capable of processing long sequences.
+  info:
+    parameters_count: 162M
+    tensor_type: F32
+    model_size: 648MB
 
 - name: mrm8488/t5-base-finetuned-summarize-news
   uri: hf://mrm8488/t5-base-finetuned-summarize-news
   description: Google's T5 base fine-tuned on News Summary dataset for summarization downstream task.
+  info:
+    parameters_count: 223M
+    tensor_type: F32
+    model_size: 892MB
+  default_parameters:
+    max_length: 200
+    min_length: 30
+    length_penalty: 2.0
+    early_stopping: true
+    no_repeat_ngram_size: 3
+    num_beams: 4
 
 - name: Falconsai/text_summarization
   uri: hf://Falconsai/text_summarization
   description: A fine-tuned variant of the T5 transformer model, designed for the task of text summarization.
+  info:
+    parameters_count: 60.5M
+    tensor_type: F32
+    model_size: 242MB
+  default_parameters:
+    max_length: 200
+    min_length: 30
+    length_penalty: 2.0
+    early_stopping: true
+    no_repeat_ngram_size: 3
+    num_beams: 4
 
 - name: mistralai/Mistral-7B-Instruct-v0.3
   uri: hf://mistralai/Mistral-7B-Instruct-v0.3
   description: Mistral-7B-Instruct-v0.3 is an instruct fine-tuned version of the Mistral-7B-v0.3.
+  info:
+    parameters_count: 7.25B
+    tensor_type: BF16
+    model_size: 14.5GB
 
 - name: gpt-4o-mini
   uri: oai://gpt-4o-mini
   description: OpenAI's GPT-4o-mini model.
+  default_parameters:
+    temperature: 1.0
+    top_p: 1.0
+    max_completion_tokens: 200
 
 - name: gpt-4-turbo
   uri: oai://gpt-4-turbo
   description: OpenAI's GPT-4 Turbo model.
+  default_parameters:
+    temperature: 1.0
+    top_p: 1.0
+    max_completion_tokens: 200
 
 - name: open-mistral-7b
   uri: mistral://open-mistral-7b
   description: Mistral's 7B model.
+  default_parameters:
+    temperature: 0.7
+    top_p: 1.0
+    max_completion_tokens: 200
 
 - name: mistralai/Mistral-7B-Instruct-v0.2
   uri: llamafile://mistralai/Mistral-7B-Instruct-v0.2
   description: A llamafile package of Mistral's 7B Instruct model.
+  info:
+    parameters_count: 7.24B
+    tensor_type: BF16
+    model_size: 14.5GB
+  default_parameters:
+    temperature: 0.7
+    top_p: 1.0
+    max_completion_tokens: 200