ml-intern / fix_space_error_report.json
bep40's picture
Update error report with DeepSeek V4 Flash routing fix details
7d99409 verified
Raw
History Blame Contribute Delete
1.61 kB
{
"status": "rebuilding_with_deepseek_flash_fix",
"space": "bep40/ml-intern",
"space_url": "https://huggingface.co/spaces/bep40/ml-intern",
"root_cause": "DeepSeek V4 Flash model IDs were routed incorrectly: (1) GLM_52_MODEL_ID was set to 'openai/deepseek/deepseek-v4-flash' which the OpenRouter routing patch sends to OpenRouter, but OpenRouter model IDs do not include the 'openai/' prefix -- causing a 404 model-not-found. (2) The routing prefix check startswith('deepseek/') never matches 'deepseek-ai/' model IDs, so 'deepseek-ai/DeepSeek-V4-Flash' without a provider tag relied on HF Router auto-routing which could hit a provider lacking structured output support.",
"fix": "Changed GLM_52_MODEL_ID from 'openai/deepseek/deepseek-v4-flash' to 'deepseek-ai/DeepSeek-V4-Flash:deepinfra' (HF Router path with explicit provider tag). Also fixed the OpenRouter routing patch to strip the 'openai/' prefix before sending the model name to OpenRouter's API. Added ':novita' tag to DeepSeek V4 Pro for deterministic HF Router routing.",
"files_changed": ["patch_models.py"],
"model_routing": {
"deepseek-ai/DeepSeek-V4-Flash": "HF Inference Router (3 providers: novita, fireworks-ai, deepinfra) - all support tools, 1M context",
"deepseek-ai/DeepSeek-V4-Flash-0731": "HF Inference Router (5 providers) - all support tools, 1M context",
"deepseek/deepseek-v4-flash-0731": "OpenRouter - available via openrouter/deepseek/deepseek-v4-flash-0731"
},
"verification": "Space rebuild triggered. Monitor Space build logs. Test model selection for DeepSeek V4 Flash after rebuild completes."
}