synth-ai 0.2.9.dev5__py3-none-any.whl → 0.2.9.dev6__py3-none-any.whl

synth_ai/experimental/synth_oss.py

@@ -1,446 +0,0 @@
-# ruff: noqa
-'''
-Synth OSS Integration Module
-
-This module provides integration with Synth's open-source inference and training APIs
-from the monorepo learning_v2 service. All APIs are OpenAI-compatible.
-
-Learning V2 APIs available for integration via lm/:
-"""
-
-# API Configuration
-SYNTH_BACKEND_URL = ""
-
-# Learning V2 Modal Service URLs
-LEARNING_V2_URLS = {
-    "dev": "https://synth-laboratories-dev--learning-v2-service-fastapi-app.modal.run",
-    "prod": "https://synth-laboratories-prod--learning-v2-service-fastapi-app.modal.run", 
-    "main": "https://synth-laboratories--learning-v2-service-fastapi-app.modal.run"
-}
-
-# ============================================================================
-# HEALTH & STATUS APIS
-# ============================================================================
-
-HEALTH_APIS = {
-    "basic_health": {
-        "method": "GET",
-        "endpoint": "/health",
-        "description": "Basic health check",
-        "response": {"status": "healthy"}
-    },
-    "detailed_health": {
-        "method": "GET", 
-        "endpoint": "/learning/health",
-        "description": "Detailed health check including GPU function availability",
-        "response": {"status": "healthy", "components": {...}}
-    }
-}
-
-# ============================================================================
-# FILE MANAGEMENT APIS
-# ============================================================================
-
-FILE_MANAGEMENT_APIS = {
-    "upload_file": {
-        "method": "POST",
-        "endpoint": "/files",
-        "description": "Upload a file for fine-tuning (JSONL format)",
-        "request": "multipart/form-data with 'file' and 'purpose'='fine-tune'",
-        "response": {
-            "id": "file-abc123",
-            "object": "file", 
-            "bytes": 1234,
-            "created_at": 1638360000,
-            "filename": "data.jsonl",
-            "purpose": "fine-tune"
-        }
-    },
-    "list_files": {
-        "method": "GET",
-        "endpoint": "/files",
-        "description": "List all uploaded files",
-        "params": {"limit": "optional"},
-        "response": {"object": "list", "data": ["file_objects"]}
-    },
-    "get_file": {
-        "method": "GET",
-        "endpoint": "/files/{file_id}",
-        "description": "Get file metadata by ID",
-        "response": "Single file object with metadata"
-    },
-    "delete_file": {
-        "method": "DELETE",
-        "endpoint": "/files/{file_id}",
-        "description": "Delete a file",
-        "response": {"id": "file-abc123", "object": "file", "deleted": True}
-    },
-    "get_file_content": {
-        "method": "GET",
-        "endpoint": "/files/{file_id}/content",
-        "description": "Download raw file content",
-        "response": "Raw file content stream"
-    }
-}
-
-# ============================================================================
-# TRAINING/FINE-TUNING APIS  
-# ============================================================================
-
-TRAINING_APIS = {
-    "create_training_job": {
-        "method": "POST",
-        "endpoint": "/fine_tuning/jobs",
-        "description": "Create a fine-tuning job",
-        "request": {
-            "model": "Qwen/Qwen3-0.5B",
-            "training_file": "file-abc123", 
-            "training_type": "sft",  # or "dpo"
-            "hyperparameters": {...},
-            "suffix": "optional"
-        },
-        "response": {
-            "object": "fine_tuning.job",
-            "id": "ftjob-xyz789",
-            "model": "...",
-            "status": "validating_files",
-            "training_file": "file-abc123",
-            "hyperparameters": {...}
-        }
-    },
-    "list_training_jobs": {
-        "method": "GET",
-        "endpoint": "/fine_tuning/jobs", 
-        "description": "List all training jobs",
-        "response": {"object": "list", "data": ["job_objects"]}
-    },
-    "get_training_job": {
-        "method": "GET",
-        "endpoint": "/fine_tuning/jobs/{job_id}",
-        "description": "Get training job status",
-        "response": {
-            "object": "fine_tuning.job",
-            "id": "ftjob-xyz789",
-            "status": "running",  # or "completed", "failed", "cancelled"
-            "fine_tuned_model": "ft:model:suffix"  # when completed
-        }
-    },
-    "cancel_training_job": {
-        "method": "POST",
-        "endpoint": "/fine_tuning/jobs/{job_id}/cancel",
-        "description": "Cancel a running training job",
-        "response": {"object": "fine_tuning.job", "id": "...", "status": "cancelled"}
-    },
-    "get_training_events": {
-        "method": "GET", 
-        "endpoint": "/fine_tuning/jobs/{job_id}/events",
-        "description": "Get training logs/events",
-        "response": {
-            "object": "list",
-            "data": [{
-                "object": "fine_tuning.job.event",
-                "level": "info",
-                "message": "Training started",
-                "created_at": 1638360000
-            }]
-        }
-    }
-}
-
-# ============================================================================
-# INFERENCE APIS
-# ============================================================================
-
-INFERENCE_APIS = {
-    "chat_completions": {
-        "method": "POST",
-        "endpoint": "/chat/completions", 
-        "description": "OpenAI-compatible chat completions for base and fine-tuned models",
-        "request": {
-            "model": "Qwen/Qwen3-0.5B",  # or "ft:Qwen/Qwen3-0.5B:suffix"
-            "messages": [{"role": "user", "content": "Hello"}],
-            "temperature": 0.7,
-            "max_tokens": 100,
-            "top_p": 1.0,
-            "stream": False,  # Set to True for streaming
-            "tools": [],  # For tool calling
-            "tool_choice": "auto"
-        },
-        "response": {
-            "id": "chatcmpl-123",
-            "object": "chat.completion",
-            "created": 1638360000,
-            "model": "Qwen/Qwen3-0.5B",
-            "choices": [{
-                "index": 0,
-                "message": {
-                    "role": "assistant",
-                    "content": "Hello! How can I help you?",
-                    "tool_calls": []  # If tools were used
-                },
-                "finish_reason": "stop"
-            }],
-            "usage": {
-                "prompt_tokens": 10,
-                "completion_tokens": 20,
-                "total_tokens": 30
-            }
-        },
-        "streaming": "Server-sent events with data: {...} format when stream=True"
-    }
-}
-
-# ============================================================================  
-# MODEL MANAGEMENT APIS
-# ============================================================================
-
-MODEL_APIS = {
-    "list_models": {
-        "method": "GET",
-        "endpoint": "/models",
-        "description": "List all available models (base and fine-tuned)",
-        "response": {
-            "object": "list",
-            "data": [{
-                "id": "Qwen/Qwen3-0.5B",
-                "object": "model", 
-                "created": 1638360000,
-                "owned_by": "learning_v2"
-            }]
-        }
-    },
-    "delete_model": {
-        "method": "DELETE",
-        "endpoint": "/models/{model_id}",
-        "description": "Delete a fine-tuned model",
-        "response": {"id": "ft:model:suffix", "object": "model", "deleted": True}
-    }
-}
-
-# ============================================================================
-# SUPPORTED MODELS
-# ============================================================================
-
-SUPPORTED_MODELS = {
-    "base_models": [
-        # Qwen 3 family
-        "Qwen/Qwen3-0.6B",
-        "Qwen/Qwen3-1.8B",
-        "Qwen/Qwen3-8B",
-        "Qwen/Qwen3-14B",
-        "Qwen/Qwen3-32B",
-        # Qwen 2.5 family
-        "Qwen/Qwen2.5-0.5B-Instruct",
-        "Qwen/Qwen2.5-1.5B-Instruct",
-        "Qwen/Qwen2.5-3B-Instruct",
-        "Qwen/Qwen2.5-7B-Instruct",
-        "Qwen/Qwen2.5-14B-Instruct",
-        "Qwen/Qwen2.5-32B-Instruct",
-        "Qwen/Qwen2.5-72B-Instruct",
-        # OLMo 2 family
-        "allenai/OLMo-2-0425-1B-Instruct",
-        "allenai/OLMo-2-1124-7B-Instruct",
-        "allenai/OLMo-2-1124-13B-Instruct"
-    ],
-    "training_types": ["sft", "dpo"],
-    "gpu_types": ["A10G", "L40S", "A100", "H100"],
-    "features": [
-        "Tool calling",
-        "Streaming responses", 
-        "Fine-tuning",
-        "Multi-GPU training",
-        "JSONL data format",
-        "OpenAI compatibility"
-    ]
-}
-
-# ============================================================================
-# INTEGRATION PLAN – Synth OSS
-# ==========================================================================
-"""
-GPU & Resource Selection
-------------------------
-Synth OSS decides the GPU based on the `ModelFamily` definition:
-• Each `ModelConfig` lists `inference_gpus` and `training_gpus`.
-• The API’s `InferenceRouter` calls `_select_gpu_for_model`, which chooses the **first recommended GPU** returned by `get_model_gpu_recommendations` (usually the `default_inference_gpu`).
-• By default the server picks the first recommended GPU, **but** we can request
-another GPU type via a custom header that the server *can* opt to honor:
-
-    X-GPU-Preference: L40S  # or A10G, A100, H100
-
-The current dev deployment already forwards this header to `InferenceRouter`,
-so adding it makes the GPU configurable without breaking existing behaviour.
-
-`InferenceConfig` therefore gets a new optional field:
-
-```python
-class InferenceConfig(BaseModel):
-    stream: bool = False
-    gpu_preference: Optional[str] = None  # "A10G", "L40S", "A100", "H100"
-    # ...future knobs (temperature, max_tokens, etc.)
-```
-
-LM will include `gpu_preference` as that header when `backend="synth"`. If the
-header is omitted or the value is not valid for the chosen model, the server
-falls back to its default selection. This keeps the API forward-compatible and
-provides explicit GPU control when supported.
-
-Only two parts of synth-ai need to change for Synth OSS inference:
-
-1.  LM() class (synth_ai.lm)
-2.  The async respond(...) coroutine on that class
-
-Extend LM with backend="synth"; when selected, issue POST requests to
-`${LEARNING_V2_URL}/chat/completions`, supporting both streaming and
-non-streaming modes and returning the same dict structure as today.
-
-Everything else (file upload, fine-tuning, model listing) lives in the
-`synth_ai.learning` package and does NOT affect LM:
-
-synth_ai/learning/
-    ├─ files.py
-    ├─ training.py
-    ├─ models.py
-    ├─ client.py
-    └─ types.py
-
-Warm-up flow
-~~~~~~~~~~~~
-`learning_v2` exposes `POST /warmup/{model_id}` and `GET /warmup/status/{model_id}`
-(via the Render proxy).  We can exploit that to reduce first-token latency.
-
-LM API addition:
-
-```python
-async def warmup(self, model: str | None = None, gpu_preference: str | None = None) -> dict:
-    """Pre-spin the container & load weights for *model* on the requested GPU.
-    Returns the JSON response from /warmup.  If *model* is None we warm-up
-    `self.model`.
-    """
-```
-
-Implementation sketch (backend == "synth")
-------------------------------------------
-1.  Determine `model_id = model or self.model`.
-2.  Build headers:
-    ```python
-    headers = {}
-    if gpu_preference:
-        headers["X-GPU-Preference"] = gpu_preference
-    ```
-3.  `POST  f"{url}/warmup/{model_id}"`.
-4.  Optionally call `GET /warmup/status/{model_id}` in a loop until
-    `status == "ready"` (exponential backoff) – or expose a separate
-    `await LM.warmup_status(model)` helper.
-
-The method is a *no-op* for the default (OpenAI) backend so existing code keeps
-working.
-
-'''
-
-
-"""
-LEARNING_v2 server-side changes required to honor `X-GPU-Preference`
-====================================================================
-1.  **Add header extraction in API layer**
-    • File: `backend/app/services/learning_v2/modal_service/api_openai_complete.py`
-    • In `@app.post("/chat/completions")` add:
-        ```python
-        gpu_pref = req.headers.get("X-GPU-Preference")
-        ```
-    • Pass `gpu_pref` to `inference_router.route_completion(...)` and
-      `route_streaming_completion(...)`.
-
-2.  **Propagate preference through `InferenceRouter`**
-    • Update signatures of `route_completion`, `route_streaming_completion` and
-      `_select_gpu_for_model` to accept `gpu_preference: Optional[str] = None`.
-    • Inside `_select_gpu_for_model`:
-        ```python
-        if gpu_preference and gpu_preference in valid_gpu_types_for_model:
-            return gpu_preference
-        ```
-      falling back to current auto-selection logic otherwise.
-
-3.  **Validate against model config**
-    • Retrieve `model_cfg.inference_gpus` and refuse unsupported choices with
-      `HTTP 400` ("GPU type not supported for this model").
-
-4.  **Expose preference in usage / logging**
-    • Include the chosen GPU in the response’s `system_fingerprint` or add a
-      field `selected_gpu` so clients can confirm the assignment.
-    • Log header + selected GPU for observability.
-
-5.  **(Optional) Accept preference in JSON body**
-    • If we want parity with LM’s `InferenceConfig`, extend
-      `ChatCompletionRequest` with an optional `gpu_preference` field and treat
-      it as a fallback if the header is missing.
-
-6.  **Update automated tests**
-    • Modify `tests/dev/learning_v2/test_multi_gpu_inference_v2.py` to set the
-      header and assert that the correct GPU function (`*_L40S_generate`, etc.)
-      was invoked (can mock `get_gpu_function`).
-
-7.  **Documentation**
-    • Add the header to `learning_v2_api_structure.txt` and README so external
-      callers know how to request a GPU.
-
-These changes are backwards-compatible: calls without the header behave exactly
-as before.
-"""
-
-"""
-Ran tool
-
-Ran tool
-
-Read file: /Users/joshuapurtell/Documents/GitHub/monorepo/backend/app/services/learning/unified_ft_service/api.py
-
-Ran tool
-Read file: /Users/joshuapurtell/Documents/GitHub/monorepo/backend/app/services/learning/unified_ft_service/api.py
-
-Ran tool
-
-Warm-up today:
-
-• `POST /warmup/{model_name}` in `unified_ft_service/api.py` kicks off an async worker that:
-  1. Calls `get_model_routing(model_name)` → this falls back to `choose_gpu(model_id, req_gpu=None)`  
-     so no GPU can be requested.
-  2. Fires the routed Modal function once (“dummy request”) to cache weights.
-
-Because the handler never accepts a GPU argument and never reads `X-GPU-Preference`, the server always warms the model with its default GPU.
-
-What has to change
-==================
-
-Server-side (learning_v2)
-1. Add an optional header or query field  
-   `X-GPU-Preference` **or** `?gpu=A100` to `POST /warmup/{model_id}`.
-2. Pass the value into `choose_gpu(model_id, req_gpu)` inside
-   `warmup_worker()`.  The helper already validates the choice, so no other
-   change is required.
-3. Expose the selected GPU in the JSON response for visibility.
-
-Client-side (synth-ai LM)
-```python
-async def warmup(
-    self,
-    model: str | None = None,
-    gpu_preference: str | None = None,
-    wait_until_ready: bool = False,
-) -> dict:
-    mdl = model or self.model
-    headers = {}
-    if gpu_preference:
-        headers["X-GPU-Preference"] = gpu_preference
-    resp = await _client.post(f"{url}/warmup/{mdl}", headers=headers)
-    if wait_until_ready:
-        while resp.json()["status"] != "warmed":
-            await asyncio.sleep(2)
-            resp = await _client.get(f"{url}/warmup/status/{mdl}")
-    return resp.json()
-```
-
-So: **the existing endpoint does not yet support GPU selection; we need to add
-the small change above on the `learning_v2` side and then LM.warmup can request
-specific GPUs.**
-"""

synth_ai/install_sqld.sh

@@ -1,40 +0,0 @@
-#!/bin/bash
-# Install sqld binary for Synth AI
-
-set -e
-
-SQLD_VERSION="v0.26.2"
-OS=$(uname -s | tr '[:upper:]' '[:lower:]')
-ARCH=$(uname -m)
-
-# Map architecture names
-case "$ARCH" in
-    x86_64) ARCH="x86_64" ;;
-    aarch64|arm64) ARCH="aarch64" ;;
-    *) echo "Unsupported architecture: $ARCH"; exit 1 ;;
-esac
-
-# Construct download URL
-URL="https://github.com/tursodatabase/libsql/releases/download/libsql-server-${SQLD_VERSION}/sqld-${OS}-${ARCH}.tar.xz"
-
-echo "📥 Downloading sqld ${SQLD_VERSION} for ${OS}-${ARCH}..."
-
-# Download and extract
-TMP_DIR=$(mktemp -d)
-cd "$TMP_DIR"
-curl -L -o sqld.tar.xz "$URL"
-tar -xf sqld.tar.xz
-
-# Install to user's local bin
-mkdir -p ~/.local/bin
-mv sqld ~/.local/bin/
-chmod +x ~/.local/bin/sqld
-
-# Clean up
-cd -
-rm -rf "$TMP_DIR"
-
-echo "✅ sqld installed to ~/.local/bin/sqld"
-echo ""
-echo "🔧 Add ~/.local/bin to your PATH if needed:"
-echo "   export PATH=\"\$HOME/.local/bin:\$PATH\""

synth_ai/learning/offline/providers.py

@@ -1,7 +0,0 @@
-## Gemini
-
-
-## OpenAI
-
-
-## Synth

synth_ai/learning/prompts/banking77_injection_eval.py

@@ -1,168 +0,0 @@
-"""
-Banking77 in-context injection evals (async, not tests)
-
-Samples a handful of Banking77 prompts and evaluates multiple override
-contexts in parallel, printing simple accuracy for each.
-
-Usage
-- Keys in .env (GROQ_API_KEY, etc.)
-- Run: uv run -q python -m synth_ai.learning.prompts.banking77_injection_eval
-  Optional env:
-    - N_SAMPLES=20 (default)
-    - MODEL=openai/gpt-oss-20b (default)
-    - VENDOR=groq (default)
-"""
-
-from __future__ import annotations
-
-import asyncio
-import os
-import random
-from typing import Any
-
-from datasets import load_dataset
-from dotenv import load_dotenv
-from synth_ai.lm.core.main_v3 import LM, build_messages
-from synth_ai.lm.overrides import LMOverridesContext
-
-
-async def classify_one(lm: LM, text: str, label_names: list[str]) -> str:
-    labels_joined = ", ".join(label_names)
-    system_message = (
-        "You are an intent classifier for the Banking77 dataset. "
-        "Given a customer message, respond with exactly one label from the list. "
-        "Return only the label text with no extra words.\n\n"
-        f"Valid labels: {labels_joined}"
-    )
-    user_message = f"Message: {text}\nLabel:"
-    messages = build_messages(system_message, user_message, images_bytes=None, model_name=lm.model)
-    resp = await lm.respond_async(messages=messages)
-    return (resp.raw_response or "").strip()
-
-
-def choose_label(pred: str, label_names: list[str]) -> str:
-    norm_pred = pred.strip().lower()
-    label_lookup = {ln.lower(): ln for ln in label_names}
-    mapped = label_lookup.get(norm_pred)
-    if mapped is not None:
-        return mapped
-
-    # Fallback: choose the label with the highest naive token overlap
-    def score(cand: str) -> int:
-        c = cand.lower()
-        return sum(1 for w in c.split() if w in norm_pred)
-
-    return max(label_names, key=score)
-
-
-async def eval_context(
-    lm: LM,
-    items: list[tuple[str, str]],
-    label_names: list[str],
-    ctx_name: str,
-    specs: list[dict[str, Any]],
-) -> tuple[str, int, int]:
-    correct = 0
-    with LMOverridesContext(specs):
-        tasks = [classify_one(lm, text, label_names) for text, _ in items]
-        results = await asyncio.gather(*tasks, return_exceptions=True)
-    for (text, gold), pred in zip(items, results, strict=False):
-        if isinstance(pred, Exception):
-            # Treat exceptions as incorrect
-            continue
-        mapped = choose_label(pred, label_names)
-        correct += int(mapped == gold)
-    return (ctx_name, correct, len(items))
-
-
-async def main() -> None:
-    load_dotenv()
-
-    n = int(os.getenv("N_SAMPLES", "20"))
-    model = os.getenv("MODEL", "openai/gpt-oss-20b")
-    vendor = os.getenv("VENDOR", "groq")
-
-    lm = LM(model=model, vendor=vendor, temperature=0.0)
-
-    print("Loading Banking77 dataset (split='test')...")
-    ds = load_dataset("banking77", split="test")
-    label_names: list[str] = ds.features["label"].names  # type: ignore
-
-    idxs = random.sample(range(len(ds)), k=min(n, len(ds)))
-    items = [
-        (ds[i]["text"], label_names[int(ds[i]["label"])])  # (text, gold_label)
-        for i in idxs
-    ]
-
-    # Define a few override contexts to compare
-    contexts: list[dict[str, Any]] = [
-        {
-            "name": "baseline (no overrides)",
-            "overrides": [],
-        },
-        {
-            "name": "nonsense prompt injection (expected worse)",
-            "overrides": [
-                {
-                    "match": {"contains": "", "role": "user"},
-                    "injection_rules": [
-                        # Heavily corrupt user text by replacing vowels
-                        {"find": "a", "replace": "x"},
-                        {"find": "e", "replace": "x"},
-                        {"find": "i", "replace": "x"},
-                        {"find": "o", "replace": "x"},
-                        {"find": "u", "replace": "x"},
-                        {"find": "A", "replace": "X"},
-                        {"find": "E", "replace": "X"},
-                        {"find": "I", "replace": "X"},
-                        {"find": "O", "replace": "X"},
-                        {"find": "U", "replace": "X"},
-                    ],
-                }
-            ],
-        },
-        {
-            "name": "injection: atm->ATM, txn->transaction",
-            "overrides": [
-                {
-                    "match": {"contains": "atm", "role": "user"},
-                    "injection_rules": [
-                        {"find": "atm", "replace": "ATM"},
-                        {"find": "txn", "replace": "transaction"},
-                    ],
-                }
-            ],
-        },
-        {
-            "name": "params: temperature=0.0",
-            "overrides": [
-                {"match": {"contains": ""}, "params": {"temperature": 0.0}},
-            ],
-        },
-        {
-            "name": "model override: 20b->120b",
-            "overrides": [
-                {"match": {"contains": ""}, "params": {"model": "openai/gpt-oss-120b"}},
-            ],
-        },
-    ]
-
-    print(f"\nEvaluating {len(contexts)} contexts on {len(items)} Banking77 samples (async)...")
-
-    # Evaluate each context sequentially but batched (each context classifies in parallel)
-    results: list[tuple[str, int, int]] = []
-    for ctx in contexts:
-        name = ctx["name"]
-        specs = ctx["overrides"]
-        print(f"Evaluating: {name} ...")
-        res = await eval_context(lm, items, label_names, name, specs)
-        results.append(res)
-
-    print("\nResults:")
-    for name, correct, total in results:
-        acc = correct / total if total else 0.0
-        print(f"- {name}: {correct}/{total} correct ({acc:.2%})")
-
-
-if __name__ == "__main__":
-    asyncio.run(main())