ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'nightly' into feature/canvas-lab

Browse source
This commit is contained in:
Roland Tannous 2026-02-24 10:08:13 +00:00
commit 3f34996288
23 changed files with 907 additions and 661 deletions
Download patch file Download diff file Expand all files Collapse all files

19
studio/backend/core/export/export.py
View file

@ -2,6 +2,7 @@
"""
Export backend - handles model exporting in various formats
"""
import json
import logging
import os
from pathlib import Path
@ -200,6 +201,18 @@ class ExportBackend:
logger.error(traceback.format_exc())
return False, f"Failed to load checkpoint: {str(e)}"
def _write_export_metadata(self, save_directory: str):
"""Write export_metadata.json with base model info for Chat page discovery."""
try:
base_model = get_base_model_from_lora(self.current_checkpoint) if self.current_checkpoint else None
metadata = {"base_model": base_model}
metadata_path = os.path.join(save_directory, "export_metadata.json")
with open(metadata_path, "w") as f:
json.dump(metadata, f, indent=2)
logger.info(f"Wrote export metadata to {metadata_path}")
except Exception as e:
logger.warning(f"Could not write export metadata: {e}")
def export_merged_model(self,
save_directory: str,
format_type: str = "16-bit (FP16)",
@ -244,6 +257,9 @@ class ExportBackend:
self.current_tokenizer,
save_method=save_method
)
# Write export metadata so the Chat page can identify the base model
self._write_export_metadata(save_directory)
logger.info(f"Model saved successfully to {save_directory}")
# Push to hub if requested
@ -297,6 +313,9 @@ class ExportBackend:
self.current_model.save_pretrained(save_directory)
self.current_tokenizer.save_pretrained(save_directory)
# Write export metadata so the Chat page can identify the base model
self._write_export_metadata(save_directory)
logger.info(f"Model saved successfully to {save_directory}")
# Push to hub if requested

13
studio/backend/core/inference/inference.py
View file

@ -6,8 +6,10 @@ from unsloth.chat_templates import get_chat_template
from transformers import TextStreamer
from peft import PeftModel, PeftModelForCausalLM
import json
import sys
import torch
from pathlib import Path
from typing import Optional, Union, Generator, Tuple
from utils.models import ModelConfig, get_base_model_from_lora
from utils.paths import is_model_cached
@ -112,7 +114,18 @@ class InferenceBackend:
# In that case, load the real processor from the base model.
from transformers import ProcessorMixin
if not (isinstance(processor, ProcessorMixin) or hasattr(processor, "image_processor")):
# For LoRA adapters, use the base model. For local merged exports,
# read export_metadata.json to find the original base model.
processor_source = config.base_model if config.is_lora else config.identifier
if not config.is_lora and config.is_local:
_meta_path = Path(config.path) / "export_metadata.json"
try:
if _meta_path.exists():
_meta = json.loads(_meta_path.read_text())
if _meta.get("base_model"):
processor_source = _meta["base_model"]
except Exception:
pass
logger.warning(
f"FastVisionModel returned {type(processor).__name__} (no image_processor) "
f"for '{model_name}' — loading proper processor from '{processor_source}'"

57
studio/backend/core/training/trainer.py
View file

@ -343,7 +343,8 @@ class UnslothTrainer:
custom_format_mapping: dict = None,
subset: str = None,
train_split: str = "train",
eval_split: str = None) -> Optional[tuple]:
eval_split: str = None,
eval_steps: float = 0.00) -> Optional[tuple]:
"""
Load and prepare dataset for training.
@ -358,6 +359,7 @@ class UnslothTrainer:
dataset = None
eval_dataset = None
has_separate_eval_source = False # True if eval comes from a separate HF split
eval_enabled = eval_steps is not None and eval_steps > 0
if local_datasets:
# Load local datasets
@ -410,23 +412,26 @@ class UnslothTrainer:
print(f"Loaded dataset from Hugging Face: {dataset_source}\n")
# Resolve eval split from a separate HF split (explicit or auto-detected)
if eval_split:
# Explicit eval split provided - load it directly
print(f"Loading explicit eval split: '{eval_split}'\n")
eval_load_kwargs = {"path": dataset_source, "split": eval_split}
if subset:
eval_load_kwargs["name"] = subset
eval_dataset = load_dataset(**eval_load_kwargs)
has_separate_eval_source = True
print(f"Loaded eval split '{eval_split}' with {len(eval_dataset)} rows\n")
else:
# Auto-detect eval split from HF (returns a separate dataset, or None)
eval_dataset = self._auto_detect_eval_split_from_hf(
dataset_source=dataset_source,
subset=subset,
)
if eval_dataset is not None:
if eval_enabled:
if eval_split:
# Explicit eval split provided - load it directly
print(f"Loading explicit eval split: '{eval_split}'\n")
eval_load_kwargs = {"path": dataset_source, "split": eval_split}
if subset:
eval_load_kwargs["name"] = subset
eval_dataset = load_dataset(**eval_load_kwargs)
has_separate_eval_source = True
print(f"Loaded eval split '{eval_split}' with {len(eval_dataset)} rows\n")
else:
# Auto-detect eval split from HF (returns a separate dataset, or None)
eval_dataset = self._auto_detect_eval_split_from_hf(
dataset_source=dataset_source,
subset=subset,
)
if eval_dataset is not None:
has_separate_eval_source = True
else:
print("Eval disabled (eval_steps <= 0), skipping eval split detection\n")
if dataset is None:
raise ValueError("No dataset provided")
@ -472,7 +477,7 @@ class UnslothTrainer:
)
eval_dataset = eval_info["dataset"]
print(f"Eval dataset formatted successfully\n")
elif not has_separate_eval_source:
elif eval_enabled and not has_separate_eval_source:
# No separate eval source — split the already-formatted dataset
formatted_dataset = dataset_info["dataset"]
split_result = self._resolve_eval_split_from_dataset(formatted_dataset)
@ -543,7 +548,7 @@ class UnslothTrainer:
def start_training(self,
dataset: Dataset,
eval_dataset: Dataset = None,
eval_steps: float = 0.01,
eval_steps: float = 0.00,
output_dir: str = "./outputs",
num_epochs: int = 3,
learning_rate: float = 5e-5,
@ -743,12 +748,16 @@ class UnslothTrainer:
# ========== EVAL CONFIGURATION ==========
eval_dataset = training_args.get('eval_dataset', None)
eval_steps_val = training_args.get('eval_steps', 0.01)
eval_steps_val = training_args.get('eval_steps', 0.00)
if eval_dataset is not None:
config_args["eval_strategy"] = "steps"
config_args["eval_steps"] = eval_steps_val
print(f"Evaluation enabled: eval_steps={eval_steps_val} (fraction of total steps)\n")
print(f"Eval dataset: {len(eval_dataset)} rows\n")
if eval_steps_val > 0:
config_args["eval_strategy"] = "steps"
config_args["eval_steps"] = eval_steps_val
print(f"✅ Evaluation enabled: eval_steps={eval_steps_val} (fraction of total steps)\n")
print(f"Eval dataset: {len(eval_dataset)} rows\n")
else:
print(f"⚠️ Eval dataset provided but eval_steps={eval_steps_val} (disabled)\n")
print("To enable evaluation, set eval_steps > 0.0\n")
else:
print("No eval dataset — evaluation disabled\n")

7
studio/backend/core/training/training.py
View file

@ -115,7 +115,7 @@ class TrainingBackend:
subset: str = None,
train_split: str = "train",
eval_split: str = None,
eval_steps: float = 0.01,
eval_steps: float = 0.00,
is_dataset_multimodal: bool = False) -> bool:
"""
Start training.
@ -223,6 +223,7 @@ class TrainingBackend:
subset=subset,
train_split=train_split,
eval_split=eval_split,
eval_steps=eval_steps,
)
# Unpack: load_and_format_dataset returns (dataset, eval_dataset)
@ -232,10 +233,6 @@ class TrainingBackend:
dataset = dataset_result
eval_dataset = None
# If user set eval_steps to 0, disable evaluation entirely
if eval_steps is not None and float(eval_steps) <= 0:
eval_dataset = None
# Track whether eval is enabled for status reporting
self.eval_enabled = eval_dataset is not None

6
studio/backend/models/models.py
View file

@ -57,10 +57,12 @@ class ModelDetails(BaseModel):
class LoRAInfo(BaseModel):
"""LoRA adapter information"""
"""LoRA adapter or exported model information"""
display_name: str = Field(..., description="Display name for the LoRA")
adapter_path: str = Field(..., description="Path to the LoRA adapter")
adapter_path: str = Field(..., description="Path to the LoRA adapter or exported model")
base_model: Optional[str] = Field(None, description="Base model identifier")
source: Optional[str] = Field(None, description="'training' or 'exported'")
export_type: Optional[str] = Field(None, description="'lora' or 'merged' (for exports)")
class LoRAScanResponse(BaseModel):

2
studio/backend/models/training.py
View file

@ -21,7 +21,7 @@ class TrainingStartRequest(BaseModel):
subset: Optional[str] = None
train_split: Optional[str] = Field("train", description="Training split name")
eval_split: Optional[str] = Field(None, description="Eval split name. None = auto-detect")
eval_steps: float = Field(0.01, description="Fraction of total steps between evals (0-1)")
eval_steps: float = Field(0.00, description="Fraction of total steps between evals (0-1)")
@model_validator(mode="before")
@classmethod

42
studio/backend/routes/models.py
View file

@ -18,6 +18,7 @@ from auth.authentication import get_current_subject
try:
from utils.models import (
scan_trained_loras,
scan_exported_models,
load_model_defaults,
get_base_model_from_lora,
is_vision_model,
@ -32,6 +33,7 @@ except ImportError:
sys.path.insert(0, str(parent_backend))
from utils.models import (
scan_trained_loras,
scan_exported_models,
load_model_defaults,
get_base_model_from_lora,
is_vision_model,
@ -289,35 +291,45 @@ async def get_model_config(
@router.get("/loras")
async def scan_loras(
outputs_dir: str = Query(default="./outputs", description="Directory to scan for LoRA adapters"),
exports_dir: str = Query(default="./exports", description="Directory to scan for exported models"),
current_subject: str = Depends(get_current_subject),
):
"""
Scan for trained LoRA adapters in the outputs directory.
This endpoint wraps the backend scan_trained_loras function.
Scan for trained LoRA adapters and exported models.
Returns both training outputs (from outputs_dir) and exported models
(from exports_dir) in a single list, distinguished by source field.
"""
try:
# Call backend scan function
trained_loras = scan_trained_loras(outputs_dir=outputs_dir)
# Convert to LoRAInfo objects
lora_list = []
# Scan training outputs
trained_loras = scan_trained_loras(outputs_dir=outputs_dir)
for display_name, adapter_path in trained_loras:
# Get base model if available
base_model = get_base_model_from_lora(adapter_path)
lora_info = LoRAInfo(
lora_list.append(LoRAInfo(
display_name=display_name,
adapter_path=adapter_path,
base_model=base_model
)
lora_list.append(lora_info)
base_model=base_model,
source="training",
))
# Scan exported models (merged, LoRA, base — skips GGUF)
exported = scan_exported_models(exports_dir=exports_dir)
for display_name, model_path, export_type, base_model in exported:
lora_list.append(LoRAInfo(
display_name=display_name,
adapter_path=model_path,
base_model=base_model,
source="exported",
export_type=export_type,
))
return LoRAScanResponse(
loras=lora_list,
outputs_dir=outputs_dir
)
except Exception as e:
logger.error(f"Error scanning LoRAs: {e}", exc_info=True)
raise HTTPException(

2
studio/backend/utils/models/__init__.py
View file

@ -5,6 +5,7 @@ from .model_config import (
ModelConfig,
is_vision_model,
scan_trained_loras,
scan_exported_models,
load_model_defaults,
get_base_model_from_lora,
load_model_config,
@ -17,6 +18,7 @@ __all__ = [
'ModelConfig',
'is_vision_model',
'scan_trained_loras',
'scan_exported_models',
'load_model_defaults',
'get_base_model_from_lora',
'load_model_config',

84
studio/backend/utils/models/model_config.py
View file

@ -465,6 +465,90 @@ def scan_trained_loras(outputs_dir: str = "./outputs") -> List[Tuple[str, str]]:
logger.error(f"Error scanning outputs folder: {e}")
return []
def scan_exported_models(exports_dir: str = "./exports") -> List[Tuple[str, str, str, Optional[str]]]:
"""
Scan exports folder for exported models (merged, LoRA, base).
Skips GGUF-only exports (not loadable by Unsloth inference backend).
The exports directory is two levels deep: {run}/{checkpoint}/
Returns:
List of tuples: [(display_name, model_path, export_type, base_model), ...]
export_type: "lora" | "merged"
"""
results = []
exports_path = Path(exports_dir)
if not exports_path.exists():
return results
try:
for run_dir in exports_path.iterdir():
if not run_dir.is_dir():
continue
for checkpoint_dir in run_dir.iterdir():
if not checkpoint_dir.is_dir():
continue
adapter_config = checkpoint_dir / "adapter_config.json"
config_file = checkpoint_dir / "config.json"
has_weights = (
any(checkpoint_dir.glob("*.safetensors"))
or any(checkpoint_dir.glob("*.bin"))
)
has_gguf = any(checkpoint_dir.glob("*.gguf"))
base_model = None
export_type = None
if adapter_config.exists():
export_type = "lora"
try:
cfg = json.loads(adapter_config.read_text())
base_model = cfg.get("base_model_name_or_path")
except Exception:
pass
elif config_file.exists() and has_weights:
export_type = "merged"
# Read base model from export_metadata.json (written at export time)
export_meta = checkpoint_dir / "export_metadata.json"
try:
if export_meta.exists():
meta = json.loads(export_meta.read_text())
base_model = meta.get("base_model")
except Exception:
pass
elif has_gguf:
# GGUF-only — not loadable by current inference backend
continue
else:
continue
# Fallback: read base model from the original training run's
# adapter_config.json in ./outputs/{run_name}/
if not base_model:
outputs_adapter_cfg = Path("./outputs") / run_dir.name / "adapter_config.json"
try:
if outputs_adapter_cfg.exists():
cfg = json.loads(outputs_adapter_cfg.read_text())
base_model = cfg.get("base_model_name_or_path")
except Exception:
pass
display_name = f"{run_dir.name} / {checkpoint_dir.name}"
model_path = str(checkpoint_dir)
results.append((display_name, model_path, export_type, base_model))
logger.debug(f"Found exported model: {display_name} ({export_type})")
results.sort(key=lambda x: Path(x[1]).stat().st_mtime, reverse=True)
logger.info(f"Found {len(results)} exported models in {exports_dir}")
return results
except Exception as e:
logger.error(f"Error scanning exports folder: {e}")
return []
def get_base_model_from_lora(lora_path: str) -> Optional[str]:
"""
Read the base model name from a LoRA adapter's config.

15
studio/frontend/src/components/assistant-ui/model-selector.tsx
View file

@ -62,7 +62,7 @@ function ModelSelectorTrigger({
className={cn(
"flex items-center gap-2 transition-colors",
variant === "outline" &&
"rounded-full border border-border/60 hover:bg-accent",
"rounded-full border border-border/60 hover:bg-accent",
variant === "ghost" && "rounded-md hover:bg-accent",
variant === "muted" && "rounded-md bg-muted hover:bg-muted/80",
size === "sm" && "h-8 px-3 text-xs",
@ -183,9 +183,20 @@ export function ModelSelector({
all.set(model.id, model);
}
for (const lora of loraModels) {
// Strip "/ suffix" from display name (e.g. "foo_123/foo" → "foo_123")
const displayName = lora.name.includes("/")
? lora.name.split("/")[0].trim()
: lora.name;
// Show type tag instead of base model name
const isExported = lora.source === "exported";
const isMerged = lora.exportType === "merged";
const tag = isExported
? isMerged ? "Merged · Exported" : "LoRA"
: "LoRA";
all.set(lora.id, {
...lora,
description: lora.baseModel || lora.description,
name: displayName,
description: tag,
});
}
return all;

29
studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
View file

@ -365,15 +365,26 @@ export function LoraModelPicker({
<div key={baseModel}>
{index > 0 ? <div className="my-1" /> : null}
<ListLabel>{baseModel}</ListLabel>
{adapters.map((adapter) => (
<ModelRow
key={adapter.id}
label={adapter.name}
meta="LoRA"
selected={value === adapter.id}
onClick={() => onSelect(adapter.id, { source: "lora", isLora: true })}
/>
))}
{adapters.map((adapter) => {
const isExported = adapter.source === "exported";
const isMerged = adapter.exportType === "merged";
const tag = isExported
? isMerged ? "Merged" : "LoRA"
: "LoRA";
const meta = isExported ? `${tag} · Exported` : tag;
return (
<ModelRow
key={adapter.id}
label={adapter.name}
meta={meta}
selected={value === adapter.id}
onClick={() => onSelect(adapter.id, {
source: isExported ? "exported" : "lora",
isLora: !isMerged,
})}
/>
);
})}
</div>
))
)}

4
studio/frontend/src/components/assistant-ui/model-selector/types.ts
View file

@ -10,10 +10,12 @@ export interface ModelOption {
export interface LoraModelOption extends ModelOption {
baseModel?: string;
updatedAt?: number;
source?: "training" | "exported";
exportType?: "lora" | "merged";
}
export interface ModelSelectorChangeMeta {
source: "hub" | "lora";
source: "hub" | "lora" | "exported";
isLora: boolean;
}

2
studio/frontend/src/config/training.ts
View file

@ -103,7 +103,7 @@ export const DEFAULT_HYPERPARAMS = {
warmupSteps: 5,
maxSteps: 0,
saveSteps: 0,
evalSteps: 0.01,
evalSteps: 0.00,
packing: false,
trainOnCompletions: false,
gradientCheckpointing: "unsloth" as const,

2
studio/frontend/src/features/chat/chat-page.tsx
View file

@ -410,6 +410,8 @@ export function ChatPage(): ReactElement {
name: lora.name,
baseModel: lora.baseModel,
updatedAt: lora.updatedAt,
source: lora.source,
exportType: lora.exportType,
})),
[lorasFromStore],
);

4
studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
View file

@ -68,6 +68,8 @@ function toLoraSummary(lora: {
display_name: string;
adapter_path: string;
base_model?: string | null;
source?: "training" | "exported" | null;
export_type?: "lora" | "merged" | null;
}): ChatLoraSummary {
const idTail = lora.adapter_path.split("/").filter(Boolean).at(-1) ?? "";
const updatedAt =
@ -78,6 +80,8 @@ function toLoraSummary(lora: {
name: stripTrailingEpoch(lora.display_name),
baseModel: lora.base_model || "Unknown base model",
updatedAt,
source: lora.source ?? undefined,
exportType: lora.export_type ?? undefined,
};
}

2
studio/frontend/src/features/chat/types/api.ts
View file

@ -14,6 +14,8 @@ export interface BackendLoraInfo {
display_name: string;
adapter_path: string;
base_model?: string | null;
source?: "training" | "exported" | null;
export_type?: "lora" | "merged" | null;
}
export interface ListLorasResponse {

2
studio/frontend/src/features/chat/types/runtime.ts
View file

@ -33,4 +33,6 @@ export interface ChatLoraSummary {
name: string;
baseModel: string;
updatedAt?: number;
source?: "training" | "exported";
exportType?: "lora" | "merged";
}

1302
studio/frontend/src/features/studio/sections/params-section.tsx
View file

File diff suppressed because it is too large Load diff

2
studio/frontend/src/features/studio/sections/progress-section.tsx
View file

@ -237,6 +237,7 @@ export function ProgressSection(): ReactElement {
onClick={() => {
setStopRequested(true);
setStopDialogOpen(false);
useTrainingRuntimeStore.getState().setStopRequested(true);
void stopTrainingRun(false).then((ok) => {
if (!ok) setStopRequested(false);
});
@ -248,6 +249,7 @@ export function ProgressSection(): ReactElement {
onClick={() => {
setStopRequested(true);
setStopDialogOpen(false);
useTrainingRuntimeStore.getState().setStopRequested(true);
void stopTrainingRun(true).then((ok) => {
if (!ok) setStopRequested(false);
});

9
studio/frontend/src/features/studio/studio-page.tsx
View file

@ -45,11 +45,16 @@ export function StudioPage(): ReactElement {
const dialogInitial = useDatasetPreviewDialogStore((s) => s.initialData);
const closeDialog = useDatasetPreviewDialogStore((s) => s.close);
const stopRequested = useTrainingRuntimeStore((state) => state.stopRequested);
const canGoBack =
showTrainingView &&
!isTrainingRunning &&
!isHydratingRuntime &&
(runtimePhase === "stopped" || runtimePhase === "error" || runtimePhase === "completed" || runtimePhase === "idle");
(stopRequested ||
(!isTrainingRunning &&
(runtimePhase === "stopped" ||
runtimePhase === "error" ||
runtimePhase === "completed" ||
runtimePhase === "idle")));
const tourEnabled = hasHydratedRuntime && !isHydratingRuntime;
const isConfigTour = !showTrainingView;
const tourSteps = showTrainingView ? studioTrainingTourSteps : studioTourSteps;

78
studio/frontend/src/features/studio/training-start-overlay.tsx
View file

@ -1,9 +1,23 @@
import {
AlertDialog,
AlertDialogAction,
AlertDialogCancel,
AlertDialogContent,
AlertDialogDescription,
AlertDialogFooter,
AlertDialogHeader,
AlertDialogTitle,
} from "@/components/ui/alert-dialog";
import { Button } from "@/components/ui/button";
import {
AnimatedSpan,
Terminal,
TypingAnimation,
} from "@/components/ui/terminal"
import type { ReactElement } from "react"
} from "@/components/ui/terminal";
import { useTrainingActions, useTrainingRuntimeStore } from "@/features/training";
import { Cancel01Icon } from "@hugeicons/core-free-icons";
import { HugeiconsIcon } from "@hugeicons/react";
import { useEffect, useState, type ReactElement } from "react";
type TrainingStartOverlayProps = {
message: string
@ -14,18 +28,65 @@ export function TrainingStartOverlay({
message,
currentStep,
}: TrainingStartOverlayProps): ReactElement {
const { stopTrainingRun } = useTrainingActions();
const isStarting = useTrainingRuntimeStore((s) => s.isStarting);
const [cancelDialogOpen, setCancelDialogOpen] = useState(false);
const [cancelRequested, setCancelRequested] = useState(false);
useEffect(() => {
if (!isStarting) {
setCancelRequested(false);
}
}, [isStarting]);
return (
<div className="pointer-events-none absolute inset-0 z-30 flex items-center justify-center rounded-2xl bg-background/45 backdrop-blur-[1px]">
<div className="flex w-[860px] max-w-[calc(100%-2rem)] flex-col items-center gap-4">
<div className="pointer-events-auto relative flex w-[860px] max-w-[calc(100%-2rem)] flex-col items-center gap-4">
<img
src="/Sloth emojis/large sloth wave.png"
alt="Unsloth mascot"
className="size-24 animate-bounce object-contain"
/>
<Terminal
className="w-full min-h-[390px] rounded-2xl px-7 py-6 text-left"
startOnView={false}
>
<div className="relative w-full">
<AlertDialog open={cancelDialogOpen} onOpenChange={setCancelDialogOpen}>
<Button
variant="ghost"
size="icon"
className="absolute right-3 top-3 z-10 size-7 cursor-pointer rounded-full text-muted-foreground/60 hover:bg-destructive/10 hover:text-destructive"
onClick={() => setCancelDialogOpen(true)}
disabled={cancelRequested}
>
<HugeiconsIcon icon={Cancel01Icon} className="size-3.5" />
</Button>
<AlertDialogContent overlayClassName="bg-background/40 supports-backdrop-filter:backdrop-blur-[1px]">
<AlertDialogHeader>
<AlertDialogTitle>Cancel Training</AlertDialogTitle>
<AlertDialogDescription>
Do you want to cancel the current training run?
</AlertDialogDescription>
</AlertDialogHeader>
<AlertDialogFooter>
<AlertDialogCancel>Continue Training</AlertDialogCancel>
<AlertDialogAction
variant="destructive"
onClick={() => {
setCancelRequested(true);
setCancelDialogOpen(false);
useTrainingRuntimeStore.getState().setStopRequested(true);
void stopTrainingRun(false).then((ok) => {
if (!ok) setCancelRequested(false);
});
}}
>
Cancel Training
</AlertDialogAction>
</AlertDialogFooter>
</AlertDialogContent>
</AlertDialog>
<Terminal
className="w-full min-h-[390px] rounded-2xl px-7 py-6 text-left"
startOnView={false}
>
<TypingAnimation
duration={36}
className="bg-gradient-to-r from-emerald-300 via-lime-300 to-teal-300 bg-clip-text font-semibold text-transparent"
@ -51,7 +112,8 @@ O^O/ \\_/ \\
<AnimatedSpan className="mt-2 text-muted-foreground">
{`> ${message || "starting training..."} | waiting for first step... (${currentStep})`}
</AnimatedSpan>
</Terminal>
</Terminal>
</div>
</div>
</div>
)

5
studio/frontend/src/features/training/stores/training-runtime-store.ts
View file

@ -37,6 +37,7 @@ const initialState: TrainingRuntimeState = {
gradNormHistory: [],
evalLossHistory: [],
resetGeneration: 0,
stopRequested: false,
};
function sortSeries(points: TrainingSeriesPoint[]): TrainingSeriesPoint[] {
@ -110,6 +111,7 @@ function applyMetricHistoryFromStatus(payload: TrainingStatusResponse): {
export const useTrainingRuntimeStore = create<TrainingRuntimeStore>()((set) => ({
...initialState,
setStopRequested: (value) => set({ stopRequested: value }),
setHydrating: (value) => set({ isHydrating: value }),
setHasHydrated: (value) => set({ hasHydrated: value }),
setStarting: (value) => set({ isStarting: value }),
@ -173,12 +175,15 @@ export const useTrainingRuntimeStore = create<TrainingRuntimeStore>()((set) => (
const detailLoss = payload.details?.loss;
const detailLr = payload.details?.learning_rate;
const detailEpoch = payload.details?.epoch;
const stopRequested =
payload.is_training_running ? state.stopRequested : false;
return {
...state,
jobId: payload.job_id || state.jobId,
phase: payload.phase,
isTrainingRunning: payload.is_training_running,
stopRequested,
evalEnabled: payload.eval_enabled ?? state.evalEnabled,
message: payload.message,
error: payload.error,

2
studio/frontend/src/features/training/types/runtime.ts
View file

@ -95,9 +95,11 @@ export interface TrainingRuntimeState {
gradNormHistory: TrainingSeriesPoint[];
evalLossHistory: TrainingSeriesPoint[];
resetGeneration: number;
stopRequested: boolean;
}
export interface TrainingRuntimeActions {
setStopRequested: (value: boolean) => void;
setHydrating: (value: boolean) => void;
setHasHydrated: (value: boolean) => void;
setStarting: (value: boolean) => void;