feat(inference): add custom model support

Author: fpagny

pages/managed-inference/reference-content/supported-models.mdx

@@ -28,7 +28,7 @@ Managed Inference supports multiple AI models either from:
 | Provider | Model string | Documentation | License |
 |-----------------|-----------------|-----------------|-----------------|
 | Meta        | `llama-3.3-70b-instruct`  | [Details](https://www.scaleway.com/en/docs/managed-inference/reference-content/llama-3.3-70b-instruct/) | [Llama 3.3 Community](https://www.llama.com/llama3_3/license/) |
-| Meta        | `llama-3.1-8b-instruct`  | [Details](https://www.scaleway.com/en/docs/managed-inference/reference-content/llama-3.1-8b-instruct/) | [HF](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) |
+| Meta        | `llama-3.1-8b-instruct`  | [Details](https://www.scaleway.com/en/docs/managed-inference/reference-content/llama-3.1-8b-instruct/) | [Llama 3.1 Community](https://llama.meta.com/llama3_1/license/) |
 
 ### Vision models
 
@@ -42,11 +42,35 @@ Managed Inference supports multiple AI models either from:
 
 ### Prerequesites
 
+<Message type="tip">
+  To begin with custom models deployment, we recommend you start with existing variation of models supported in the Scaleway Catalog. As an example, you can deploy a [quantized version (4 bits) of Llama 3.3](https://huggingface.co/unsloth/Llama-3.3-70B-Instruct-bnb-4bit). If you want to then deploy a fine-tuned version of Llama 3.3, you can ensure the file structure you provide matches this example before creating your deployment.
+</Message>
+
 To deploy a model by providing its URL on Hugging Face, you need to:
 - Have access to this model with your Hugging Face credentials (if the model is "Gated", you specifically need to ask access from your Hugging Face account). Note that your Hugging Face credentials will not be stored, but we still recommend you to create [read or fine-grained access tokens](https://huggingface.co/docs/hub/security-tokens) for this purpose. 
 
-### Additional consideration
+The model files need to include:
+- a `config.json` file containing:
+  - `architectures` array. See [supported models architectures](#supported-models-architecture) for exact list of supported values.
+  - `max_position_embeddings`
+- model weigths in [`.safetensors`](https://huggingface.co/docs/safetensors/index) format
+- a chat template either in:
+  - `tokenizer_config.json` file as `chat_template` field
+  - `chat_template.json` file as `chat_template` field
+
+For security reasons, models containing arbitrary code execution such as [`pickle`](https://docs.python.org/3/library/pickle.html) format are not supported. 
+
+### Custom model lifecycle
+
+Currently, custom model deployments are considered to be valid for a long term, and we will ensure any updatse or changes to Managed Inference will not impact existing deployments.
+In case of breaking changes, leading to some custom models not being supported anymore, we will notify you at least 3 months beforehand.
+
+### License
+
+- When deploying custom models, you remain responsible for complying with any License requirements from the model provider, as you would do by running the model on a custom provisioned GPU.
+
+### Supported models architecture
+
+Custom Models Deployments currently support the following models architecture: `AquilaModel`, `AquilaForCausalLM`, `ArcticForCausalLM`, `BaiChuanForCausalLM`, `BaichuanForCausalLM`, `BloomForCausalLM`, `CohereForCausalLM`, `Cohere2ForCausalLM`, `DbrxForCausalLM`, `DeciLMForCausalLM`, `DeepseekForCausalLM`, `DeepseekV2ForCausalLM`, `DeepseekV3ForCausalLM`, `ExaoneForCausalLM`, `FalconForCausalLM`, `Fairseq2LlamaForCausalLM`, `GemmaForCausalLM`, `Gemma2ForCausalLM`, `GlmForCausalLM`, `GPT2LMHeadModel`, `GPTBigCodeForCausalLM`, `GPTJForCausalLM`, `GPTNeoXForCausalLM`, `GraniteForCausalLM`, `GraniteMoeForCausalLM`, `GritLM`, `InternLMForCausalLM`, `InternLM2ForCausalLM`, `InternLM2VEForCausalLM`, `InternLM3ForCausalLM`, `JAISLMHeadModel`, `JambaForCausalLM`, `LlamaForCausalLM`, `LLaMAForCausalLM`, `MambaForCausalLM`, `FalconMambaForCausalLM`, `MiniCPMForCausalLM`, `MiniCPM3ForCausalLM`, `MistralForCausalLM`, `MixtralForCausalLM`, `QuantMixtralForCausalLM`, `MptForCausalLM`, `MPTForCausalLM`, `NemotronForCausalLM`, `OlmoForCausalLM`, `Olmo2ForCausalLM`, `OlmoeForCausalLM`, `OPTForCausalLM`, `OrionForCausalLM`, `PersimmonForCausalLM`, `PhiForCausalLM`, `Phi3ForCausalLM`, `Phi3SmallForCausalLM`, `PhiMoEForCausalLM`, `Qwen2ForCausalLM`, `Qwen2MoeForCausalLM`, `RWForCausalLM`, `StableLMEpochForCausalLM`, `StableLmForCausalLM`, `Starcoder2ForCausalLM`, `SolarForCausalLM`, `TeleChat2ForCausalLM`, `XverseForCausalLM`, `BartModel`, `BartForConditionalGeneration`, `Florence2ForConditionalGeneration`, `BertModel`, `RobertaModel`, `RobertaForMaskedLM`, `XLMRobertaModel`, `DeciLMForCausalLM`, `Gemma2Model`, `GlmForCausalLM`, `GritLM`, `InternLM2ForRewardModel`, `JambaForSequenceClassification`, `LlamaModel`, `MistralModel`, `Phi3ForCausalLM`, `Qwen2Model`, `Qwen2ForCausalLM`, `Qwen2ForRewardModel`, `Qwen2ForProcessRewardModel`, `TeleChat2ForCausalLM`, `LlavaNextForConditionalGeneration`, `Phi3VForCausalLM`, `Qwen2VLForConditionalGeneration`, `Qwen2ForSequenceClassification`, `BertForSequenceClassification`, `RobertaForSequenceClassification`, `XLMRobertaForSequenceClassification`, `AriaForConditionalGeneration`, `Blip2ForConditionalGeneration`, `ChameleonForConditionalGeneration`, `ChatGLMModel`, `ChatGLMForConditionalGeneration`, `DeepseekVLV2ForCausalLM`, `FuyuForCausalLM`, `H2OVLChatModel`, `InternVLChatModel`, `Idefics3ForConditionalGeneration`, `LlavaForConditionalGeneration`, `LlavaNextForConditionalGeneration`, `LlavaNextVideoForConditionalGeneration`, `LlavaOnevisionForConditionalGeneration`, `MantisForConditionalGeneration`, `MiniCPMO`, `MiniCPMV`, `MolmoForCausalLM`, `NVLM_D`, `PaliGemmaForConditionalGeneration`, `Phi3VForCausalLM`, `PixtralForConditionalGeneration`, `QWenLMHeadModel`, `Qwen2VLForConditionalGeneration`, `Qwen2_5_VLForConditionalGeneration`, `Qwen2AudioForConditionalGeneration`, `UltravoxModel`, `MllamaForConditionalGeneration`, `WhisperForConditionalGeneration`, `EAGLEModel`, `MedusaModel`, `MLPSpeculatorPreTrainedModel`
 
-- When deploying custom models, you remain responsible for complying with any License requirements from the model provider.
-- We currently