Summary

When importing mistralai/Mistral-7B-Instruct-v0.3 or mistralai/Mistral-Nemo-Instruct-2407 from HuggingFace, OME classifies both as EMBEDDING instead of TEXT_GENERATION. Chat/completions requests against the resulting endpoint return:

400 Bad Request: importedModel does not support any of: [TextToText, ImageTextToText]

Both models have "architectures": ["MistralForCausalLM"] in their HF config.json and should be TEXT_GENERATION.

Root Cause

Two code locations interact to produce the bug:

1. pkg/hfutil/modelconfig/mistral.go — GetArchitecture() fallback

func (c *MistralConfig) GetArchitecture() string {
    if len(c.Architectures) > 0 {
        return c.Architectures[0]
    }
    return "MistralModel"   // ← dangerous fallback
}

If Architectures is empty (e.g. JSON parsing fails, field missing, or struct mismatch), the method silently returns "MistralModel".

2. pkg/modelagent/config_parser.go — determineModelCapabilitiesFromHF()

if strings.Contains(strings.ToLower(architecture), "embedding") ||
    strings.Contains(strings.ToLower(architecture), "sentence") ||
    strings.Contains(strings.ToLower(modelType), "bert") ||
    // Special case for known embedding models
    (strings.Contains(strings.ToLower(modelType), "mistral") &&
        strings.Contains(strings.ToLower(architecture), "mistralmodel")) {
    return append(capabilities, string(v1beta1.ModelCapabilityEmbedding))
}

When the fallback fires, modelType = "mistral" and architecture = "MistralModel" satisfy the special-case condition, and the model is classified as EMBEDDING.

The intended path for intfloat/e5-mistral-7b-instruct (a genuine embedding model) is correct: its HF config has architectures: [] or uses the base MistralModel architecture, so the fallback correctly labels it. The problem is that causal-LM models whose Architectures field fails to populate get the same treatment.

Repro

Import either of these models via the OME model-agent and check the resulting ClusterBaseModel.spec.modelCapabilities:

• mistralai/Mistral-7B-Instruct-v0.3 — architectures: ["MistralForCausalLM"] — classified as EMBEDDING ❌
• mistralai/Mistral-Nemo-Instruct-2407 — architectures: ["MistralForCausalLM"] — classified as EMBEDDING ❌
• intfloat/e5-mistral-7b-instruct — embedding model — classified as EMBEDDING ✅

Expected Behaviour

Proposed Fix

Change the GetArchitecture() fallback from "MistralModel" to "" so a missing/unparsed Architectures field does not accidentally satisfy the embedding special-case:

func (c *MistralConfig) GetArchitecture() string {
    if len(c.Architectures) > 0 {
        return c.Architectures[0]
    }
    return ""   // don't assume MistralModel; let caller treat as unknown
}

Alternatively, tighten the special-case check in config_parser.go to require the architecture to be exactly "MistralModel" (case-insensitive) rather than a substring match, and only when Architectures was explicitly set (not via fallback).

Additional Context

• autoSelect is false on the vllm-e5-mistral-7b-instruct runtime and the two runtimes use distinct modelArchitecture values (MistralModel vs MistralForCausalLM), so runtime auto-selection is not affected — the runtimes cannot be confused with each other.
• The misclassification only affects capability gating at the endpoint level (chat vs embedding API routing).