What happened?

Health checks fail for Bedrock cross-region inference models that are not explicitly listed in model_prices_and_context_window.json, even though normal acompletion calls to the same models work fine.

Root Cause

_update_litellm_params_for_health_check (litellm/proxy/health_check.py:260-284) strips the bedrock/ prefix from the model name before passing it to ahealth_check. After stripping, get_llm_provider can no longer identify the provider from the model name alone and falls back to checking litellm.bedrock_converse_models set. If the model is not in that set (populated from the cost map), provider identification fails.

Normal acompletion calls do NOT go through this pre-processing, so they keep the bedrock/ prefix and get_llm_provider can identify the provider by splitting on /.

Flow comparison

Normal acompletion (works):

"bedrock/jp.anthropic.claude-sonnet-4-6"
  → get_llm_provider splits on "/" → provider="bedrock" → OK

Health check (fails):

"bedrock/jp.anthropic.claude-sonnet-4-6"
  → _update_litellm_params_for_health_check strips "bedrock/"
  → "jp.anthropic.claude-sonnet-4-6"
  → get_llm_provider cannot split on "/" to find provider
  → falls back to checking bedrock_converse_models set
  → "jp.anthropic.claude-sonnet-4-6" not in set → provider unknown → FAIL

Why some cross-region models pass and others don't

The bedrock_converse_models set is populated from model_prices_and_context_window.json during litellm.__init__. Models that happen to have entries in the cost map (e.g., jp.anthropic.claude-sonnet-4-5-20250929-v1:0) are in the set and pass health checks. Models without entries (e.g., jp.anthropic.claude-sonnet-4-6) are not in the set and fail.

The pre-processing logic

The _update_litellm_params_for_health_check function (added for Issue #15807) strips bedrock/ and then removes full AWS region names (us-west-2, ap-northeast-1, etc.) from the path. However, it does not account for cross-region inference profile prefixes (us., eu., jp., etc.), which use a different format (dot-separated, not path-separated).

Possible Fixes

Option A: Preserve bedrock/ prefix in health check pre-processing - only strip the region path segment, not the provider prefix:

# Instead of stripping "bedrock/" entirely, keep it
# "bedrock/us-west-2/model" → "bedrock/model" (strip region only)
# "bedrock/jp.anthropic.claude-sonnet-4-6" → unchanged (no region to strip)

Option B: Pass custom_llm_provider explicitly to ahealth_check so it doesn't need to be inferred from the model name.

Steps to Reproduce

1. Configure a model not in the cost map with a cross-region inference profile:

model_list:
  - model_name: sonnet-4-6
    litellm_params:
      model: bedrock/jp.anthropic.claude-sonnet-4-6

1. Verify normal calls work:

# This works
response = await litellm.acompletion(
    model="bedrock/jp.anthropic.claude-sonnet-4-6",
    messages=[{"role": "user", "content": "hi"}]
)

1. Health check fails:

GET /health → model "jp.anthropic.claude-sonnet-4-6" fails

1. Compare with a model that IS in the cost map:

# This passes health checks because jp.anthropic.claude-sonnet-4-5-20250929-v1:0
# exists in model_prices_and_context_window.json
model_list:
  - model_name: sonnet-4-5
    litellm_params:
      model: bedrock/jp.anthropic.claude-sonnet-4-5-20250929-v1:0

Relevant Code

• Health check pre-processing: litellm/proxy/health_check.py _update_litellm_params_for_health_check() (line ~225)
• Provider identification: litellm/litellm_core_utils/get_llm_provider_logic.py get_llm_provider() (line ~99)
• bedrock_converse_models population: litellm/__init__.py (line ~724)

Related Issues

• #24669 — Cross-region inference pricing entries unreachable due to lookup order (same root data issue, different symptom)
• #15807 — Original issue that introduced the health check bedrock/ stripping logic

Component

Proxy

LiteLLM Version

v1.82.6