Skip to content

Profiles & models

lm15.profiles.ProviderProfile / EndpointProfile are deprecated (1.0.0rc2) and removed in 1.0.0; see the migration table in Using model profiles. The compatibility policies and the request-level hatch stay.

lm15.profiles — provider endpoint profiles (DEPRECATED) and the Responses compat resolution that survives them.

ProviderProfile / EndpointProfile / OpenAILM.from_profile / OpenAILM(profile=...) are deprecated as of 1.0.0rc2 and removed in 1.0.0 (contract changes/2026-09-11-job-handles-live-turns-profiles.md § 3). Everything they expressed has one home now:

  • an endpoint's address and wire policy: OpenAILM(compat="ollama") (a preset name supplies its own address) or compat= + base_url=;
  • a per-model policy: the chat dialect's OpenAIChatCompat.model_overrides, or per request Config.extensions["openai_responses_compat"];
  • a compat guessed from the base URL: nothing — say compat="openrouter".

What stays, in lm15.compat where it belongs: the partial-compat semantics (None inherits, "auto" is the adapter default), merge_openai_responses_compat, and the request-level hatch (:func:openai_responses_compat_from_extensions). ModelInfo and ModelRegistry were never profiles; they live in lm15.models.

OpenAIChatCompat dataclass

Partial compatibility policy for OpenAI Chat Completions-family APIs.

This class is consumed by OpenAIChatLM (lm15.providers.openai_chat); it is kept separate so profiles can describe chat-completions style endpoints without overloading OpenAIResponsesCompat.

for_model(model: str) -> 'OpenAIChatCompat'

This compat with the first matching model_overrides entry applied.

preset(name: str) -> 'OpenAIChatCompat' classmethod

The named server dialect from :data:OPENAI_CHAT_PRESETS.

Accepts the permanent spelling aliases (lm-studio, z.ai, openai_chat); raises ValueError for an unknown name.

OpenAIResponsesCompat dataclass

Partial compatibility policy for OpenAI Responses-family APIs.

None means "inherit". Non-None values override parent profiles. The value "auto" means "explicitly use adapter auto-detection".

preset(name: str) -> 'OpenAIResponsesCompat' classmethod

The named server dialect from :data:OPENAI_RESPONSES_PRESETS.

Accepts the permanent spelling aliases (responses, lm-studio, z.ai); raises ValueError for an unknown name.

ResolvedOpenAIResponsesCompat dataclass

Fully resolved OpenAI Responses compatibility policy.

InferencePricing dataclass

A model's list prices, per million tokens, and estimate.

Rates are per million tokens of each billing bucket of MAP-19: uncached input (input_tokens - cache_read_tokens - cache_write_tokens), cache reads, cache writes, output (reasoning included). An absent rate is unknown, never zero: a catalog that knows a bucket costs the same as plain input states that rate. tiers (ascending by threshold) reprice long prompts; dimensions is opaque catalog data that estimate does not read.

estimate(usage: 'Usage') -> float | None

The list-price cost of one call, or None when it is not known.

Exact or nothing (spec/types.md, InferencePricing): the result is the sum, in this order, of uncached input, cache reads, cache writes and output, each tokens * rate / 1_000_000, at the base rates or at the highest tier whose above_input_tokens the call's input_tokens exceeds. None when:

  • input_tokens or output_tokens is unknown;
  • a cache counter is unknown and its rate differs from the input rate (or either rate is unknown), so the split would change the price;
  • a bucket with tokens has no rate;
  • the call has audio tokens (audio has its own prices, which these rates do not carry).

Not covered, and never guessed: the price differences lm15 cannot see in Usage: service tiers (flex, priority, batch), regional multipliers, a cache write's longer lifetime, server-side tool fees. The number is the standard list price of the tokens reported.

ModelRegistry dataclass

from_dicts(dicts: Iterable[dict]) -> 'ModelRegistry' classmethod

Build a registry from canonical ModelInfo dicts.

Each dict is validated through serde.model_info_from_dict; the ModelInfo constructors reject junk (negative prices, empty modality names, missing id/provider) by raising ValueError/TypeError.

discover(*, group: str = 'lm15.model_catalogs') -> 'ModelRegistry' classmethod

Hydrate a registry from installed entry-point catalogs.

Each entry point in group must load to a zero-argument callable returning an iterable of canonical ModelInfo dicts (docs/model-hydration.md). Catalogs are processed sorted by entry-point name; on duplicate (provider, id) keys the FIRST occurrence wins. A catalog that raises is skipped with a warning — discovery never crashes the host application.

Hydrated data is ADVISORY metadata only: it must never change what build_request produces.

EndpointProfile dataclass

DEPRECATED (1.0.0rc2; removed in 1.0.0). See the module docstring.

ProviderProfile dataclass

DEPRECATED (1.0.0rc2; removed in 1.0.0). See the module docstring.

openai_responses_compat_from_extensions(extensions: JsonObject | None) -> OpenAIResponsesCompat | None

Read request-level OpenAI Responses compat from Config.extensions.

Supported shapes:

{"openai_responses_compat": {...}}
{"openai_compat": {...}}                 # backwards-friendly alias
{"compat": {"openai_responses": {...}}}
{"compat": {"openai": {...}}}            # generic OpenAI alias

The per-request hatch: the one place a Responses-dialect policy can differ per model after profiles are gone. Normal configuration is OpenAILM(compat=...).

resolve_openai_responses_compat(*, base_url: str, model: str, profile: ProviderProfile | None, request_extensions: JsonObject | None, base: OpenAIResponsesCompat | None = None) -> ResolvedOpenAIResponsesCompat

Resolve effective OpenAI Responses compatibility policy.

Layering:

bound compat (``OpenAILM(compat=...)``), else the base URL default
< endpoint compat            (deprecated with profiles)
< model compat               (deprecated with profiles)
< request extension override

The base-URL default is a guess about the server and is deprecated with the profiles: when it would pick anything but OpenAI's own policy it warns, naming the compat= spelling that says the same thing explicitly.

lm15.models — Optional model metadata and registry utilities.

The canonical inference request remains Request(model="..."). ModelInfo and ModelRegistry are optional helpers for model discovery, validation, routing, and cost estimation. Model capabilities are endpoint-specific (inference today) so new endpoint families can be described later without breaking the inference model. Fine-tune PROVENANCE stays describable through ModelOrigin (type, base_model) — lm15 does inference with tuned models, not training.

PricingTier dataclass

The rates of a call whose prompt is longer than above_input_tokens.

A tier reprices the WHOLE call, not only the tokens past the threshold: a call with Usage.input_tokens > above_input_tokens pays every rate from the highest such tier (OpenAI: "Long context: >272K input tokens"; Anthropic Haiku 5.5: "a request whose prompt is over 100,000 tokens pays higher prices", cache reads and writes included). input_tokens is the whole prompt (MAP-19), which is what both providers count. A rate absent from a tier is unknown at that tier; it is never taken from the base rates.

InferencePricing dataclass

A model's list prices, per million tokens, and estimate.

Rates are per million tokens of each billing bucket of MAP-19: uncached input (input_tokens - cache_read_tokens - cache_write_tokens), cache reads, cache writes, output (reasoning included). An absent rate is unknown, never zero: a catalog that knows a bucket costs the same as plain input states that rate. tiers (ascending by threshold) reprice long prompts; dimensions is opaque catalog data that estimate does not read.

estimate(usage: 'Usage') -> float | None

The list-price cost of one call, or None when it is not known.

Exact or nothing (spec/types.md, InferencePricing): the result is the sum, in this order, of uncached input, cache reads, cache writes and output, each tokens * rate / 1_000_000, at the base rates or at the highest tier whose above_input_tokens the call's input_tokens exceeds. None when:

  • input_tokens or output_tokens is unknown;
  • a cache counter is unknown and its rate differs from the input rate (or either rate is unknown), so the split would change the price;
  • a bucket with tokens has no rate;
  • the call has audio tokens (audio has its own prices, which these rates do not carry).

Not covered, and never guessed: the price differences lm15 cannot see in Usage: service tiers (flex, priority, batch), regional multipliers, a cache write's longer lifetime, server-side tool fees. The number is the standard list price of the tokens reported.

ModelRegistry dataclass

from_dicts(dicts: Iterable[dict]) -> 'ModelRegistry' classmethod

Build a registry from canonical ModelInfo dicts.

Each dict is validated through serde.model_info_from_dict; the ModelInfo constructors reject junk (negative prices, empty modality names, missing id/provider) by raising ValueError/TypeError.

discover(*, group: str = 'lm15.model_catalogs') -> 'ModelRegistry' classmethod

Hydrate a registry from installed entry-point catalogs.

Each entry point in group must load to a zero-argument callable returning an iterable of canonical ModelInfo dicts (docs/model-hydration.md). Catalogs are processed sorted by entry-point name; on duplicate (provider, id) keys the FIRST occurrence wins. A catalog that raises is skipped with a warning — discovery never crashes the host application.

Hydrated data is ADVISORY metadata only: it must never change what build_request produces.

Compatibility policies

For a worked Chat Completions example, see Connect an unlisted OpenAI-compatible server. Preset strings must be recognized names; use a typed policy object for an unlisted server's format differences.

Partial compatibility policy for OpenAI Chat Completions-family APIs.

This class is consumed by OpenAIChatLM (lm15.providers.openai_chat); it is kept separate so profiles can describe chat-completions style endpoints without overloading OpenAIResponsesCompat.

for_model(model: str) -> 'OpenAIChatCompat'

This compat with the first matching model_overrides entry applied.

preset(name: str) -> 'OpenAIChatCompat' classmethod

The named server dialect from :data:OPENAI_CHAT_PRESETS.

Accepts the permanent spelling aliases (lm-studio, z.ai, openai_chat); raises ValueError for an unknown name.

Partial compatibility policy for OpenAI Responses-family APIs.

None means "inherit". Non-None values override parent profiles. The value "auto" means "explicitly use adapter auto-detection".

preset(name: str) -> 'OpenAIResponsesCompat' classmethod

The named server dialect from :data:OPENAI_RESPONSES_PRESETS.

Accepts the permanent spelling aliases (responses, lm-studio, z.ai); raises ValueError for an unknown name.