[CRITICAL] Frontend silently drops settings fields during model create/update — speculative decoding, cache types, multimodal paths lost #12

New issue

Closed

opened 2026-05-08 23:46:25 +00:00 by Mark · 1 comment

Mark commented

2026-05-08 23:46:25 +00:00

Owner

Summary

The frontend helper functions vllmSettingsToApi() and llamaCppSettingsToApi() in src/hooks/useModels.ts do not convert many settings fields to the snake_case API format, causing them to be silently dropped when creating or updating models.

Affected File

src/hooks/useModels.ts

Missing in `vllmSettingsToApi()`:

cpu_offload_gb / cpuOffloadGb
draft_hf_repo / draftHfRepo
draft_quantization / draftQuantization
draft_tensor_parallel_size / draftTensorParallelSize
draft_max_model_len / draftMaxModelLen
num_speculative_tokens / numSpeculativeTokens
attention_backend / attentionBackend
kv_cache_dtype / kvCacheDtype

Missing in `llamaCppSettingsToApi()`:

cpu_moe / cpuMoe
n_cpu_moe / nCpuMoe
fit_to_memory / fitToMemory
draft_gguf_file / draftGgufFile
num_speculative_tokens / numSpeculativeTokens
flash_attn / flashAttn
cache_type_k / cacheTypeK
cache_type_v / cacheTypeV
mmproj_path / mmprojPath
draft_gguf_file / draftGgufFile

Impact

Any model created or updated via the frontend UI will have these settings silently stripped. For example:

Speculative decoding configurations are lost
Multimodal mmproj path is lost
Cache type (TurboQuant) settings are lost
Flash attention settings are lost

Expected Behavior

All settings fields should be included in the conversion to snake_case API format.

## Summary The frontend helper functions `vllmSettingsToApi()` and `llamaCppSettingsToApi()` in `src/hooks/useModels.ts` do not convert many settings fields to the snake_case API format, causing them to be silently dropped when creating or updating models. ## Affected File `src/hooks/useModels.ts` ### Missing in `vllmSettingsToApi()`: - `cpu_offload_gb` / `cpuOffloadGb` - `draft_hf_repo` / `draftHfRepo` - `draft_quantization` / `draftQuantization` - `draft_tensor_parallel_size` / `draftTensorParallelSize` - `draft_max_model_len` / `draftMaxModelLen` - `num_speculative_tokens` / `numSpeculativeTokens` - `attention_backend` / `attentionBackend` - `kv_cache_dtype` / `kvCacheDtype` ### Missing in `llamaCppSettingsToApi()`: - `cpu_moe` / `cpuMoe` - `n_cpu_moe` / `nCpuMoe` - `fit_to_memory` / `fitToMemory` - `draft_gguf_file` / `draftGgufFile` - `num_speculative_tokens` / `numSpeculativeTokens` - `flash_attn` / `flashAttn` - `cache_type_k` / `cacheTypeK` - `cache_type_v` / `cacheTypeV` - `mmproj_path` / `mmprojPath` - `draft_gguf_file` / `draftGgufFile` ## Impact Any model created or updated via the frontend UI will have these settings silently stripped. For example: - Speculative decoding configurations are lost - Multimodal mmproj path is lost - Cache type (TurboQuant) settings are lost - Flash attention settings are lost ## Expected Behavior All settings fields should be included in the conversion to snake_case API format.