fix: Ensure endpoints deploy command uses the deploy id and revision id when provided (#463)

* fix: Ensure `endpoints deploy` command uses the deploy id and revision id when provided

* fix(cli): reject duplicate endpoint model revisions

Co-authored-by: Blaine Kasten <blainekasten@gmail.com>

* refactor(cli): handle model not found by exception type

Co-authored-by: Blaine Kasten <blainekasten@gmail.com>

---------

Co-authored-by: Cursor Agent <cursoragent@cursor.com>

Author: blainekasten

src/together/lib/cli/api/beta/endpoints/_utils/_resolve_model.py

@@ -1,8 +1,9 @@
 from __future__ import annotations
 
 import re
+from typing import NamedTuple
 
-from together import APIError
+from together import NotFoundError
 from together.types.beta import Model, Endpoint
 from together.lib.cli.utils.config import CLIConfigParameter
 from together.lib.cli.utils._console import console
@@ -18,9 +19,12 @@
 # Logic for resolving a model + config from a user input string
 #
 # 1. Raw model id (e.g. ml_...)
-#    → GET /configs?referenceModelId=... for the config and model path
+#    → retrieve from --project when possible; config via baseModelId / id
+#    → else GET /configs?referenceModelId=... (public / reference models)
 # 2. Full model path (projects/.../models/...)
-#    → parse the model id, then same as (1)
+#    → retrieve that model (keep it as the deploy target), resolve config via
+#      baseModelId (or the model id when it is itself a reference model).
+#      Preserve an optional /revisions/... pin on the deploy path.
 # 3. Named model (prefix/model-name)
 #    a. prefix == project slug → list private models by name, resolve config via
 #       baseModelId (path 1), but deploy the custom model path
@@ -31,7 +35,13 @@
 #   - one profile + no --config → use that profile's config
 #   - --config given → use the profile whose config matches
 # After either path, re-validate against the user's --config when provided.
-MODEL_PATH_RE = re.compile(r"^projects/([^/]+)/models/([^/]+)(?:/revisions/[^/]+)?$")
+MODEL_PATH_RE = re.compile(r"^projects/([^/]+)/models/([^/]+)(?:/revisions/([^/]+))?$")
+
+
+class ResolvedModelAndConfig(NamedTuple):
+    model: Model
+    config: Config
+    revision_id: str | None = None
 
 
 async def resolve_model(
@@ -40,24 +50,47 @@ async def resolve_model(
     *,
     config_id: str | None = None,
 ) -> Model:
-    model, _config = await resolve_model_and_config(config, model_input, config_id=config_id)
-    return model
+    resolved = await resolve_model_and_config(config, model_input, config_id=config_id)
+    return resolved.model
 
 
 async def resolve_model_and_config(
     config: CLIConfigParameter,
     model_input: str,
     *,
     config_id: str | None = None,
-) -> tuple[Model, Config]:
+) -> ResolvedModelAndConfig:
     """Resolve a deployable model and the config revision to pair with it."""
-    # 1 / 2. Full model path or raw model id → configs list by referenceModelId
+    # 2. Full model path → keep the user's model; config from its base/reference.
     path_match = MODEL_PATH_RE.match(model_input)
     if path_match:
-        _project_id, model_id = path_match.groups()
-        return await _resolve_via_configs(config, model_id, config_id=config_id, model_input=model_input)
+        project_id, model_id, revision_id = path_match.group(1), path_match.group(2), path_match.group(3)
+        return await _resolve_explicit_model(
+            config,
+            model_id=model_id,
+            project_id=project_id,
+            config_id=config_id,
+            model_input=model_input,
+            revision_id=revision_id,
+        )
 
+    # 1. Raw model id
     if "/" not in model_input:
+        if config.project_id:
+            try:
+                model = await config.client.beta.models.retrieve(id=model_input, project_id=config.project_id)
+            except NotFoundError:
+                pass
+            else:
+                reference_model_id = model.base_model_id or model.id
+                assert reference_model_id is not None
+                return await _resolve_config_for_model(
+                    config,
+                    model,
+                    reference_model_id=reference_model_id,
+                    config_id=config_id,
+                    model_input=model_input,
+                )
         return await _resolve_via_configs(config, model_input, config_id=config_id, model_input=model_input)
 
     # 3. Named model (prefix/model-name)
@@ -70,32 +103,82 @@ async def resolve_model_and_config(
         reference_model_id = model.base_model_id or model.id
         assert reference_model_id is not None
         # Config comes from the base/reference model; deploy path stays the custom model.
-        _base_model, selected_config = await _resolve_via_configs(
+        return await _resolve_config_for_model(
             config,
-            reference_model_id,
+            model,
+            reference_model_id=reference_model_id,
             config_id=config_id,
             model_input=model_input,
         )
-        return model, selected_config
 
     return await _resolve_public_model_and_config(config, model_input, config_id=config_id)
 
 
+async def _resolve_explicit_model(
+    config: CLIConfigParameter,
+    *,
+    model_id: str,
+    project_id: str,
+    config_id: str | None,
+    model_input: str,
+    revision_id: str | None = None,
+) -> ResolvedModelAndConfig:
+    """Load the user-specified model and pair it with a compatible config."""
+    try:
+        model = await config.client.beta.models.retrieve(id=model_id, project_id=project_id)
+    except NotFoundError:
+        raise ValueError(f"Model {model_input} not found.") from None
+
+    reference_model_id = model.base_model_id or model.id
+    assert reference_model_id is not None
+    return await _resolve_config_for_model(
+        config,
+        model,
+        reference_model_id=reference_model_id,
+        config_id=config_id,
+        model_input=model_input,
+        revision_id=revision_id,
+    )
+
+
+async def _resolve_config_for_model(
+    config: CLIConfigParameter,
+    model: Model,
+    *,
+    reference_model_id: str,
+    config_id: str | None,
+    model_input: str,
+    revision_id: str | None = None,
+) -> ResolvedModelAndConfig:
+    selected = resolve_config(
+        await resolve_configs(config, reference_model_id),
+        config_id,
+        model=model_input,
+    )
+    selected = validate_requested_config(selected, config_id, model=model_input)
+    return ResolvedModelAndConfig(model=model, config=selected, revision_id=revision_id)
+
+
 async def _resolve_via_configs(
     config: CLIConfigParameter,
     reference_model_id: str,
     *,
     config_id: str | None,
     model_input: str,
-) -> tuple[Model, Config]:
+) -> ResolvedModelAndConfig:
+    """Resolve a public/reference model id through the configs API.
+
+    The deploy target is the config's reference model — correct when the user
+    passed a bare reference-model id that is not retrievable under --project.
+    """
     selected = resolve_config(
         await resolve_configs(config, reference_model_id),
         config_id,
         model=model_input,
     )
     selected = validate_requested_config(selected, config_id, model=model_input)
     model = await _retrieve_model_from_reference(config, selected, model_input=model_input)
-    return model, selected
+    return ResolvedModelAndConfig(model=model, config=selected)
 
 
 async def _retrieve_model_from_reference(
@@ -108,18 +191,16 @@ async def _retrieve_model_from_reference(
     path = selected.reference_model or ""
     match = MODEL_PATH_RE.match(path)
     if match:
-        project_id, model_id = match.groups()
+        project_id, model_id = match.group(1), match.group(2)
     elif selected.reference_model_id and selected.project_id:
         project_id, model_id = selected.project_id, selected.reference_model_id
     else:
         raise ValueError(f"Config {selected.id} has no usable reference model path.")
 
     try:
         return await config.client.beta.models.retrieve(id=model_id, project_id=project_id)
-    except APIError as e:
-        if "not found" in e.message.lower():
-            raise ValueError(f"Model {model_input} not found.") from None
-        raise
+    except NotFoundError:
+        raise ValueError(f"Model {model_input} not found.") from None
 
 
 async def _find_private_model_by_name(config: CLIConfigParameter, name: str) -> Model:
@@ -236,7 +317,7 @@ async def _resolve_public_model_and_config(
     model_input: str,
     *,
     config_id: str | None = None,
-) -> tuple[Model, Config]:
+) -> ResolvedModelAndConfig:
     supported_models = await config.client.beta.models.list_supported(search=model_input)
     if not supported_models.data:
         raise ValueError(f"Model {model_input} not found.")
@@ -255,19 +336,22 @@ async def _resolve_public_model_and_config(
     match = MODEL_PATH_RE.match(profile.model or "")
     if not match:
         raise ValueError(f"Invalid model path: {profile.model}")
-    project_id, model_id = match.groups()
+    project_id, model_id, revision_id = match.group(1), match.group(2), match.group(3)
 
     selected_config = validate_requested_config(
         config_from_profile(profile),
         config_id,
         model=model_input,
     )
     model = Model.construct(id=model_id, projectId=project_id, name=public_model.name or model_id)
-    return model, selected_config
+    return ResolvedModelAndConfig(model=model, config=selected_config, revision_id=revision_id)
 
 
-def construct_model_path(model: Model) -> str:
-    return f"projects/{model.project_id}/models/{model.id}"
+def construct_model_path(model: Model, revision_id: str | None = None) -> str:
+    path = f"projects/{model.project_id}/models/{model.id}"
+    if revision_id:
+        return f"{path}/revisions/{revision_id}"
+    return path
 
 
 async def resolve_endpoint(config: CLIConfigParameter, endpoint_id_or_name: str) -> Endpoint:

src/together/lib/cli/api/beta/endpoints/ab.py

@@ -74,7 +74,8 @@ async def ab(
     )
     verify_control_receiving_traffic(endpoint, control)
 
-    resolved_model, config_value = await resolve_model_and_config(config, model, config_id=config_id)
+    resolved = await resolve_model_and_config(config, model, config_id=config_id)
+    resolved_model, config_value = resolved.model, resolved.config
 
     autoscaling = build_autoscaling(
         min_replicas=1,
@@ -103,7 +104,7 @@ async def ab(
             endpoint_id=endpoint.id,
             enable_lora=enable_lora,
             name=name,
-            model=construct_model_path(resolved_model),
+            model=construct_model_path(resolved_model, resolved.revision_id),
             config=construct_config_path(config_value),
             autoscaling=autoscaling,
         ),

src/together/lib/cli/api/beta/endpoints/deploy.py

@@ -30,6 +30,7 @@
     PlacementProfile,
 )
 from together.lib.cli.api.beta.endpoints._utils._resolve_model import (
+    MODEL_PATH_RE,
     construct_model_path,
     resolve_model_and_config,
 )
@@ -174,7 +175,18 @@ async def deploy(
     config: CLIConfigParameter,
 ) -> None:
     """Create a deployment on a new or existing dedicated inference endpoint."""
-    resolved_model, config_value = await resolve_model_and_config(config, model, config_id=config_id)
+    model_path_match = MODEL_PATH_RE.match(model)
+    if model_revision is not None and model_path_match is not None and model_path_match.group(3) is not None:
+        raise ValueError(
+            "Do not pass --model-revision when --model already includes a revision. "
+            "Specify the revision only in the fully qualified --model path."
+        )
+
+    resolved = await resolve_model_and_config(config, model, config_id=config_id)
+    resolved_model, config_value = resolved.model, resolved.config
+    # Prefer revision pin from a fully-qualified model path; fall back to the
+    # deprecated --model-revision flag.
+    resolved_revision = resolved.revision_id or model_revision
 
     autoscaling = build_autoscaling(
         min_replicas=min_replicas,
@@ -200,16 +212,18 @@ async def deploy(
     else:
         placement_value = placement.to_json()
 
+    model_path = construct_model_path(resolved_model, resolved_revision)
+
     if not config.json:
         _print_deployment_preview(
             endpoint=endpoint_name_or_id,
             deployment_name=deployment_name,
             model=resolved_model,
+            model_path=model_path,
             config_value=config_value,
             autoscaling=autoscaling,
             placement=placement_value,
             enable_lora=enable_lora,
-            model_revision=model_revision,
             traffic_weight=traffic_weight,
         )
     await assert_explicit_project_id(config)
@@ -222,11 +236,12 @@ async def deploy(
             config.client.beta.endpoints.deployments.create(
                 endpoint.id,
                 name=deployment_name,
-                model=construct_model_path(resolved_model),
+                model=model_path,
                 config=construct_config_path(config_value),
                 autoscaling=autoscaling,
                 enable_lora=enable_lora if enable_lora is not None else omit,
-                model_revision_id=model_revision or omit,
+                # Revision is already embedded in model_path when present.
+                model_revision_id=omit,
                 placement=placement_value or omit,
             ),
         )
@@ -267,11 +282,11 @@ def _print_deployment_preview(
     endpoint: str,
     deployment_name: str,
     model: Model,
+    model_path: str,
     config_value: Config,
     autoscaling: DeploymentAutoscalingParam,
     placement: Placement | None,
     enable_lora: bool | None,
-    model_revision: str | None,
     traffic_weight: float | None,
 ) -> None:
     table = Table(expand=True, show_header=False, show_edge=False, show_lines=False, box=None, pad_edge=False)
@@ -302,9 +317,6 @@ def add_row(flag: str, value: str) -> None:
         if percentile := metric.get("percentile"):
             add_row("--scaling-percentile", percentile)
 
-    if model_revision:
-        add_row("--model-revision", model_revision)
-
     if placement is not None:
         if "profile" in placement:
             add_row("--placement", placement["profile"])  # type: ignore[typeddict-item]
@@ -324,7 +336,7 @@ def add_row(flag: str, value: str) -> None:
         add_row("--enable-lora", "true" if enable_lora else "false")
     if traffic_weight is not None:
         add_row("--traffic-weight", str(traffic_weight))
-    add_row("--model", model.name)
+    add_row("--model", f"{model.name} ({model_path})")
     add_row("--config", config_value.id)  # type: ignore
 
     table.add_row("\n".join(args))

src/together/lib/cli/api/beta/endpoints/shadow.py

@@ -92,7 +92,8 @@ async def shadow(
     rate, target_qps = await resolve_rate_or_target_qps(rate, target_qps, config=config)
 
     endpoint_id = (await resolve_endpoint(config, endpoint_id_or_name)).id
-    resolved_model, config_value = await resolve_model_and_config(config, model, config_id=config_id)
+    resolved = await resolve_model_and_config(config, model, config_id=config_id)
+    resolved_model, config_value = resolved.model, resolved.config
 
     autoscaling = build_autoscaling(
         min_replicas=1,
@@ -122,7 +123,7 @@ async def shadow(
             endpoint_id=endpoint_id,
             name=name,
             enable_lora=enable_lora,
-            model=construct_model_path(resolved_model),
+            model=construct_model_path(resolved_model, resolved.revision_id),
             config=construct_config_path(config_value),
             autoscaling=autoscaling,
         ),

src/together/lib/cli/api/beta/models/create.py

@@ -30,7 +30,7 @@ async def preprompt(self, config: CLIConfig) -> None:
             for profile in model.deployment_profiles:
                 match = MODEL_PATH_RE.match(profile.model)
                 if match:
-                    _, model_id = match.groups()
+                    model_id = match.group(2)
                     self.choices.append((f"{model.name} ({profile.quantization})", model_id))