feat(databricks): add User-Agent partner attribution support

[jatorre]

Summary

Adds support for ISV partner attribution via Databricks SDK User-Agent headers, enabling proper identification and tracking of requests from partner integrations.

Problem

Databricks requires ISV partners to identify their requests via User-Agent for proper attribution and analytics, as documented in the Databricks SDK.

This is currently not configurable in LiteLLM, preventing ISV partners from meeting Databricks compliance requirements.

Solution

• Added multi-tier configuration system with clear priority hierarchy:

  1. Per-request metadata (highest priority)
  2. Model-level litellm_params
  3. Global litellm_settings
  4. Environment variables (lowest priority)

• Follows existing Vertex AI labels pattern for consistency

• Backwards compatible (all new parameters are optional)

• Works seamlessly with proxy and Responses API

Configuration Examples

Global proxy config (recommended for proxy deployments):

litellm_settings:
  databricks_partner: "my-company"
  databricks_product: "my-product"
  databricks_product_version: "1.0.0"

Per-request metadata (for dynamic scenarios):

litellm.completion(
    model="databricks/databricks-meta-llama-3-1-70b-instruct",
    messages=[{"role": "user", "content": "Hello"}],
    metadata={
        "databricks_partner": "my-company",
        "databricks_product": "my-product",
        "databricks_product_version": "1.0.0"
    }
)

Model-level configuration:

model_list:
  - model_name: databricks-llama
    litellm_params:
      model: databricks/databricks-meta-llama-3-1-70b-instruct
      databricks_partner: "my-company"
      databricks_product: "my-product"
      databricks_product_version: "1.0.0"

Environment variables (fallback):

export DATABRICKS_SDK_UPSTREAM="my-company"
export DATABRICKS_SDK_UPSTREAM_VERSION="my-product/1.0.0"

Benefits

• ✅ ISV partners can meet Databricks compliance requirements
• ✅ Better usage tracking and analytics
• ✅ Flexible configuration for different deployment scenarios
• ✅ Generally useful (not company-specific)
• ✅ Follows existing LiteLLM patterns (similar to Vertex AI labels)

Testing

• Added 6 comprehensive test cases covering:
  • Priority level 1: Per-request metadata
  • Priority level 2: Model-level params
  • Priority level 3: Global settings
  • Priority level 4: Environment variable preservation
  • No-config scenario
  • Version-only scenario
• All tests verify proper environment variable handling
• Syntax validated and backwards compatibility confirmed

Files Changed

• litellm/llms/databricks/common_utils.py - Core implementation
• litellm/llms/databricks/chat/transformation.py - Integration point
• tests/test_litellm/llms/databricks/test_databricks_common_utils.py - Test coverage

Related

This addresses the ISV integration requirements documented in Databricks SDK and enables proper partner attribution for any organization using LiteLLM as a proxy to Databricks Foundation Model APIs.

[vercel]

@jatorre is attempting to deploy a commit to the CLERKIEAI Team on Vercel.

A member of the Team first needs to authorize it.

[krrishdholakia]

global settings are at the module level - litellm.<attribute>

[krrishdholakia]

optional params will handle both per request and model level values

[krrishdholakia]

this should not be needed. optional params handles both

[krrishdholakia]

did you manually qa this? it looks wrong

[jatorre]

Sorry for the initial messy PR! I've now studied how other providers (AWS Bedrock, Azure, etc.) handle similar configurations and simplified significantly.

What Changed

Removed unnecessary complexity:

• Checked existing patterns in the codebase - no other provider uses module-level settings for provider-specific config
• We don't need tenant-level metadata, so removed that entire priority level
• Simplified from 4 priorities → 2 priorities (matching AWS Bedrock pattern)
• Reduced tests from 5 → 2 (kept only essential coverage)

Current implementation:

• 48 lines total (20 implementation + 28 tests)
• 2-priority system: optional_params → environment variables
• Follows existing LiteLLM patterns

Usage:

# Per-request or model-level (via optional_params)
litellm.completion(
    model="databricks/databricks-meta-llama-3-1-70b-instruct",
    messages=[{"role": "user", "content": "Hello"}],
    databricks_partner="carto",
    databricks_product="agentic-gis",
    databricks_product_version="1.0.0"
)