Add seeded Text LLM sampling

docs/src/content/docs/features/prompt-tools.md

@@ -2,7 +2,7 @@
 title: LLM Prompt Tools
 sidebar:
   order: 3
-lastUpdated: 2026-05-23
+lastUpdated: 2026-08-03
 ---
 
 InvokeAI includes two built-in tools that use local language models to help you write better prompts. Both tools appear as small buttons in the top-right corner of the positive prompt area and are only visible when you have a compatible model installed.
@@ -52,4 +52,6 @@ Both tools overwrite your current prompt. You can undo this change:
 
 ## Workflow Node
 
-A **Text LLM** node is also available in the workflow editor for use in automated pipelines. It accepts a prompt string and model selection as inputs and outputs the expanded text as a string.
+The workflow editor provides **Text LLM** and **Text LLM (with System Prompt Preset)** nodes for automated pipelines. Both accept a prompt, model, maximum token count, and seed, then output generated text as a string. The preset variant reads its system prompt from the System Prompts library.
+
+Using the same seed reproduces sampling when the model, prompt, settings, hardware, and software versions remain the same. Results may differ across devices or software versions. The prompt area's **Expand Prompt** tool chooses a fresh seed for each request so repeated expansions can vary.

invokeai/app/api/routers/utilities.py

@@ -20,6 +20,7 @@
 from invokeai.app.services.image_files.image_files_common import ImageFileNotFoundException
 from invokeai.app.services.model_records.model_records_base import UnknownModelException
 from invokeai.app.util.dynamicprompts import find_missing_wildcards
+from invokeai.app.util.misc import get_random_seed
 from invokeai.backend.llava_onevision_pipeline import LlavaOnevisionPipeline
 from invokeai.backend.model_manager.taxonomy import ModelType
 from invokeai.backend.text_llm_pipeline import DEFAULT_SYSTEM_PROMPT, ProgressCallback, TextLLMPipeline
@@ -166,6 +167,7 @@ def _run_expand_prompt(
             prompt=prompt,
             system_prompt=system_prompt or DEFAULT_SYSTEM_PROMPT,
             max_new_tokens=max_tokens,
+            seed=get_random_seed(),
             device=model_device,
             dtype=TorchDevice.choose_torch_dtype(),
             progress_callback=progress_callback,

invokeai/app/invocations/text_llm.py

@@ -10,6 +10,7 @@
     SystemPromptNotFoundError,
     SystemPromptRecordDTO,
 )
+from invokeai.app.util.misc import SEED_MAX
 from invokeai.backend.model_manager.taxonomy import ModelType
 from invokeai.backend.text_llm_pipeline import DEFAULT_SYSTEM_PROMPT, TextLLMPipeline
 from invokeai.backend.util.devices import TorchDevice
@@ -21,6 +22,7 @@ def _run_text_llm(
     prompt: str,
     system_prompt: str,
     max_tokens: int,
+    seed: int,
 ) -> str:
     """Shared LLM invocation body used by every text-LLM node in this module."""
     model_config = context.models.get_config(text_llm_model)
@@ -35,6 +37,7 @@ def _run_text_llm(
             prompt=prompt,
             system_prompt=system_prompt,
             max_new_tokens=max_tokens,
+            seed=seed,
             device=model_device,
             dtype=TorchDevice.choose_torch_dtype(),
         )
@@ -45,7 +48,7 @@ def _run_text_llm(
     title="Text LLM",
     tags=["llm", "text", "prompt"],
     category="llm",
-    version="1.0.0",
+    version="1.1.0",
     classification=Classification.Beta,
 )
 class TextLLMInvocation(BaseInvocation):
@@ -72,6 +75,7 @@ class TextLLMInvocation(BaseInvocation):
         le=2048,
         description="Maximum number of tokens to generate.",
     )
+    seed: int = InputField(default=0, ge=0, le=SEED_MAX, description=FieldDescriptions.seed)
 
     @torch.no_grad()
     def invoke(self, context: InvocationContext) -> StringOutput:
@@ -81,6 +85,7 @@ def invoke(self, context: InvocationContext) -> StringOutput:
             prompt=self.prompt,
             system_prompt=self.system_prompt,
             max_tokens=self.max_tokens,
+            seed=self.seed,
         )
         return StringOutput(value=output)
 
@@ -90,7 +95,7 @@ def invoke(self, context: InvocationContext) -> StringOutput:
     title="Text LLM (with System Prompt Preset)",
     tags=["llm", "text", "prompt", "preset", "template"],
     category="llm",
-    version="1.0.0",
+    version="1.1.0",
     classification=Classification.Beta,
 )
 class TextLLMWithPresetInvocation(BaseInvocation):
@@ -125,6 +130,7 @@ class TextLLMWithPresetInvocation(BaseInvocation):
         le=2048,
         description="Maximum number of tokens to generate.",
     )
+    seed: int = InputField(default=0, ge=0, le=SEED_MAX, description=FieldDescriptions.seed)
 
     def _resolve_system_prompt(self, context: InvocationContext) -> SystemPromptRecordDTO:
         """Resolve the referenced preset, enforcing the same access rules as the REST API.
@@ -169,5 +175,6 @@ def invoke(self, context: InvocationContext) -> StringOutput:
             prompt=self.prompt,
             system_prompt=record.content,
             max_tokens=self.max_tokens,
+            seed=self.seed,
         )
         return StringOutput(value=output)

invokeai/backend/text_llm_pipeline.py

@@ -4,6 +4,7 @@
 from typing import Callable
 
 import torch
+from torch.overrides import TorchFunctionMode
 from transformers import PreTrainedModel, PreTrainedTokenizerBase, TextIteratorStreamer
 
 DEFAULT_SYSTEM_PROMPT = (
@@ -29,6 +30,37 @@
 PROGRESS_EMIT_INTERVAL = 0.1
 
 
+class _SeededMultinomialMode(TorchFunctionMode):
+    """Use invocation-local generators for multinomial sampling in this thread."""
+
+    def __init__(self, seed: int):
+        self._seed = seed
+        self._generators: dict[torch.device, torch.Generator] = {}
+
+    def _get_generator(self, device: torch.device) -> torch.Generator:
+        generator_device = torch.device("cpu") if device.type == "mps" else device
+        if generator_device not in self._generators:
+            self._generators[generator_device] = torch.Generator(device=generator_device).manual_seed(self._seed)
+        return self._generators[generator_device]
+
+    def __torch_function__(self, func, types, args=(), kwargs=None):
+        kwargs = dict(kwargs or {})
+        if func is not torch.multinomial or kwargs.get("generator") is not None:
+            return func(*args, **kwargs)
+
+        probabilities = args[0] if args else kwargs["input"]
+        generator = self._get_generator(probabilities.device)
+        kwargs["generator"] = generator
+        if probabilities.device.type != "mps":
+            return func(*args, **kwargs)
+
+        if args:
+            args = (probabilities.cpu(), *args[1:])
+        else:
+            kwargs["input"] = probabilities.cpu()
+        return func(*args, **kwargs).to(probabilities.device)
+
+
 class TextLLMPipeline:
     """A wrapper for a causal language model + tokenizer for text generation."""
 
@@ -41,6 +73,7 @@ def run(
         prompt: str,
         system_prompt: str = DEFAULT_SYSTEM_PROMPT,
         max_new_tokens: int = 300,
+        seed: int = 0,
         device: torch.device = torch.device("cpu"),
         dtype: torch.dtype = torch.float16,
         progress_callback: ProgressCallback | None = None,
@@ -91,7 +124,8 @@ def run(
 
         def _generate() -> None:
             try:
-                self._model.generate(**generation_kwargs)
+                with _SeededMultinomialMode(seed):
+                    self._model.generate(**generation_kwargs)
             except BaseException as e:
                 generation_error.append(e)
                 # transformers only calls streamer.end() on the normal exit of the

invokeai/frontend/web/openapi.json

@@ -81824,6 +81824,18 @@
             "title": "Max Tokens",
             "type": "integer"
           },
+          "seed": {
+            "default": 0,
+            "description": "Seed for random number generation",
+            "field_kind": "input",
+            "input": "any",
+            "maximum": 4294967295,
+            "minimum": 0,
+            "orig_default": 0,
+            "orig_required": false,
+            "title": "Seed",
+            "type": "integer"
+          },
           "type": {
             "const": "text_llm",
             "default": "text_llm",
@@ -81836,7 +81848,7 @@
         "tags": ["llm", "text", "prompt"],
         "title": "Text LLM",
         "type": "object",
-        "version": "1.0.0",
+        "version": "1.1.0",
         "output": {
           "$ref": "#/components/schemas/StringOutput"
         }
@@ -81927,6 +81939,18 @@
             "title": "Max Tokens",
             "type": "integer"
           },
+          "seed": {
+            "default": 0,
+            "description": "Seed for random number generation",
+            "field_kind": "input",
+            "input": "any",
+            "maximum": 4294967295,
+            "minimum": 0,
+            "orig_default": 0,
+            "orig_required": false,
+            "title": "Seed",
+            "type": "integer"
+          },
           "type": {
             "const": "text_llm_with_preset",
             "default": "text_llm_with_preset",
@@ -81939,7 +81963,7 @@
         "tags": ["llm", "text", "prompt", "preset", "template"],
         "title": "Text LLM (with System Prompt Preset)",
         "type": "object",
-        "version": "1.0.0",
+        "version": "1.1.0",
         "output": {
           "$ref": "#/components/schemas/StringOutput"
         }

invokeai/frontend/web/src/services/api/schema.ts

@@ -35633,6 +35633,12 @@ export type components = {
              * @default 300
              */
             max_tokens?: number;
+            /**
+             * Seed
+             * @description Seed for random number generation
+             * @default 0
+             */
+            seed?: number;
             /**
              * type
              * @default text_llm
@@ -35694,6 +35700,12 @@ export type components = {
              * @default 300
              */
             max_tokens?: number;
+            /**
+             * Seed
+             * @description Seed for random number generation
+             * @default 0
+             */
+            seed?: number;
             /**
              * type
              * @default text_llm_with_preset

tests/app/invocations/test_text_llm_with_preset.py

@@ -26,6 +26,7 @@ def _make_invocation(prompt_id: str = "test-id") -> TextLLMWithPresetInvocation:
         system_prompt=SystemPromptField(system_prompt_id=prompt_id),
         text_llm_model=ModelIdentifierField(key="dummy", hash="x", name="dummy", base="any", type="text_llm"),
         max_tokens=50,
+        seed=123,
     )
 
 
@@ -70,6 +71,7 @@ def test_preset_node_loads_content_from_db_and_passes_to_llm() -> None:
     assert kwargs["system_prompt"] == "custom system instruction"
     assert kwargs["prompt"] == "a cat"
     assert kwargs["max_tokens"] == 50
+    assert kwargs["seed"] == 123
 
     assert result.value == "expanded"
 

tests/backend/text_llm/test_text_llm_api_models.py

@@ -1,9 +1,19 @@
 """Tests for TextLLM API request/response models and validation."""
 
+from unittest.mock import MagicMock, patch
+
 import pytest
+import torch
 from pydantic import ValidationError
 
-from invokeai.app.api.routers.utilities import ExpandPromptRequest, ExpandPromptResponse, ImageToPromptRequest
+from invokeai.app.api.dependencies import ApiDependencies
+from invokeai.app.api.routers.utilities import (
+    ExpandPromptRequest,
+    ExpandPromptResponse,
+    ImageToPromptRequest,
+    _run_expand_prompt,
+)
+from invokeai.backend.model_manager.taxonomy import ModelType
 
 
 class TestExpandPromptRequest:
@@ -52,3 +62,26 @@ def test_success_response(self):
     def test_error_response(self):
         resp = ExpandPromptResponse(expanded_prompt="", error="Model failed")
         assert resp.error == "Model failed"
+
+
+def test_expand_prompt_uses_fresh_seed() -> None:
+    model_config = MagicMock(type=ModelType.TextLLM, path="model")
+    model = MagicMock()
+    model.parameters.return_value = iter([torch.nn.Parameter(torch.zeros(1))])
+    loaded_model = MagicMock()
+    loaded_model.model_on_device.return_value.__enter__.return_value = (None, model)
+    services = MagicMock()
+    services.model_manager.store.get_model.return_value = model_config
+    services.model_manager.load.load_model.return_value = loaded_model
+
+    with (
+        patch.object(ApiDependencies, "invoker", MagicMock(services=services), create=True),
+        patch("invokeai.app.api.routers.utilities._resolve_model_path", return_value="model"),
+        patch("invokeai.app.api.routers.utilities.AutoTokenizer.from_pretrained"),
+        patch("invokeai.app.api.routers.utilities.get_random_seed", return_value=123),
+        patch("invokeai.app.api.routers.utilities.TextLLMPipeline") as pipeline_class,
+    ):
+        pipeline_class.return_value.run.return_value = "expanded"
+        assert _run_expand_prompt("cat", "model", 10, None, None, "user") == "expanded"
+
+    assert pipeline_class.return_value.run.call_args.kwargs["seed"] == 123