Feature/llm api support

configs/config_mock_openai_batch.yaml

@@ -0,0 +1,57 @@
+processors:
+  - type: llm
+    server_args:
+      model_path: Qwen/Qwen3-4B-Instruct-2507
+      tp_size: 1
+      disable_custom_all_reduce: true
+    default_sampling_params:
+      temperature: 0.1
+      top_p: 0.9
+      max_new_tokens: 1024
+      custom_params:
+        chat_template_kwargs:
+          enable_thinking: false
+    batch_provider:
+      enabled: true
+      provider: openai
+      model: gpt-4o-mini
+      max_chunk_bytes: 52428800
+      metadata_output_path: tests/output/batch_metadata.jsonl
+      credentials:
+        api_key: 
+
+loading_params:
+  datasets:
+    - path: tests/mock_data/data.jsonl
+      type: JSONL
+      output_dir: tests/output/data_openai_batch
+
+  num_shards: 1
+  shard_id: 0
+  batch_size: 64
+
+processing_params:
+  inputs:
+    - name: text
+      key: text
+
+  outputs:
+    - name: formatted_answer
+      type: llm
+      output_type: JSON
+      output_schema:
+        - question
+        - answer
+      prompt: |
+        Generate one question and its corresponding answer using the following text:
+        ```
+        {{ text }}
+        ```
+
+  remove_columns: true
+  output_schema:
+    conversations:
+      - role: "user"
+        content: "{{ formatted_answer.question }}"
+      - role: "assistant"
+        content: "{{ formatted_answer.answer }}"

configs/config_mock_openai_batch_vision.yaml

@@ -0,0 +1,47 @@
+processors:
+  - type: llm
+    server_args:
+      model_path: Qwen/Qwen3-VL-8B-Instruct
+      tp_size: 1
+      trust_remote_code: true
+    chat_template: qwen2-vl
+    default_sampling_params:
+      temperature: 0.1
+      top_p: 0.9
+      max_new_tokens: 512
+    batch_provider:
+      enabled: true
+      provider: openai
+      model: gpt-4o-mini
+      metadata_output_path: tests/output/batch_metadata.jsonl
+      credentials:
+        api_key: ""
+
+loading_params:
+  datasets:
+    - path: tests/mock_data_vision/data.jsonl
+      type: JSONL
+      output_dir: tests/output/data_openai_batch_vision
+      image_base_path: tests/mock_data_vision
+
+  num_shards: 1
+  shard_id: 0
+  batch_size: 1
+
+processing_params:
+  inputs:
+    - name: image_input
+      key: image
+      type: image
+
+  outputs:
+    - name: caption
+      type: llm
+      output_type: plain
+      prompt: |
+        Describe what you see in this image in one concise sentence.
+
+  remove_columns: false
+  output_schema:
+    image: "{{ image_input }}"
+    caption: "{{ caption }}"

src/mmirage/config/batch_provider.py

@@ -0,0 +1,78 @@
+"""Provider-agnostic batch configuration contracts.
+
+This module defines the shared configuration shape used by any future batch
+submission provider (OpenAI, Anthropic, etc.).
+"""
+
+from dataclasses import dataclass, field
+from typing import Any, Dict, Literal, Optional
+
+
+@dataclass
+class BatchRetryPolicy:
+    """Retry behavior used by provider-neutral batch submission orchestration.
+
+    Attributes:
+        max_attempts: Maximum number of submission attempts for retryable errors.
+        initial_backoff_seconds: Delay before the first retry attempt.
+        backoff_multiplier: Multiplicative factor for subsequent retry delays.
+    """
+
+    max_attempts: int = 3
+    initial_backoff_seconds: float = 2.0
+    backoff_multiplier: float = 2.0
+
+    def __post_init__(self) -> None:
+        if self.max_attempts < 1:
+            raise ValueError("max_attempts must be >= 1")
+        if self.initial_backoff_seconds < 0:
+            raise ValueError("initial_backoff_seconds must be >= 0")
+        if self.backoff_multiplier < 1:
+            raise ValueError("backoff_multiplier must be >= 1")
+
+
+@dataclass
+class BatchProviderConfig:
+    """Shared contract for provider-specific batch configuration.
+
+    Concrete provider configs should inherit from this dataclass and extend it
+    with provider-specific settings. The fields here are intentionally provider
+    neutral so chunking/submission orchestration can run through one typed path.
+
+    Attributes:
+        provider: Provider identifier (for example, "openai" or "anthropic").
+        enabled: Whether batch submission mode is enabled.
+        max_chunk_bytes: Maximum serialized request bytes per chunk.
+            Defaults to 50 MB.
+        max_requests_per_chunk: Optional hard cap on number of requests in a
+            chunk. If None, no request-count cap is enforced.
+        metadata_output_path: Path where submission metadata artifacts are saved.
+        retry_policy: Retry policy used by the shared batch layer.
+        oversized_request_policy: Handling policy when a single request exceeds
+            ``max_chunk_bytes``. ``isolate`` creates a dedicated oversized
+            chunk, while ``reject`` fails fast.
+        extras: Provider-specific knobs that do not belong in the shared fields.
+        credentials: Provider credentials required to submit chunks.
+    """
+
+    provider: str
+    enabled: bool = True
+    max_chunk_bytes: int = 50 * 1024 * 1024
+    max_requests_per_chunk: Optional[int] = None
+    metadata_output_path: str = ""
+    retry_policy: BatchRetryPolicy = field(default_factory=BatchRetryPolicy)
+    oversized_request_policy: Literal["isolate", "reject"] = "isolate"
+    extras: Dict[str, Any] = field(default_factory=dict)
+    credentials: Dict[str, str] = field(default_factory=dict)
+
+    def __post_init__(self) -> None:
+        self.provider = self.provider.strip().lower()
+
+        if not self.provider:
+            raise ValueError("provider must be a non-empty string")
+        if self.max_chunk_bytes < 1:
+            raise ValueError("max_chunk_bytes must be >= 1")
+        if self.max_requests_per_chunk is not None and self.max_requests_per_chunk < 1:
+            raise ValueError("max_requests_per_chunk must be >= 1 when provided")
+        if self.oversized_request_policy not in {"isolate", "reject"}:
+            raise ValueError("oversized_request_policy must be either 'isolate' or 'reject'")

src/mmirage/config/openai_batch.py

@@ -0,0 +1,44 @@
+"""OpenAI-specific batch configuration."""
+
+from dataclasses import dataclass, field
+from typing import Any, Dict, Literal, Optional
+
+from mmirage.config.batch_provider import BatchProviderConfig
+
+
+@dataclass
+class OpenAIBatchConfig(BatchProviderConfig):
+    """OpenAI Batch API configuration.
+
+    Attributes:
+        provider: Fixed provider identifier for OpenAI.
+        model: Model name used in each chat completion request body.
+        batch_endpoint: Target endpoint used by OpenAI batch jobs.
+        completion_window: OpenAI completion window value.
+        base_url: Optional base URL, useful for API-compatible gateways.
+        metadata: Metadata sent on batch creation.
+    """
+
+    provider: str = "openai"
+    model: str = "gpt-4.1-mini"
+    batch_endpoint: str = "/v1/chat/completions"
+    completion_window: Literal["24h"] = "24h"
+    base_url: Optional[str] = None
+    metadata: Dict[str, str] = field(default_factory=dict)
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+
+        if not self.model.strip():
+            raise ValueError("model must be a non-empty string")
+        if not self.batch_endpoint.startswith("/"):
+            raise ValueError("batch_endpoint must start with '/'")
+
+        # Mirror OpenAI-specific fields into generic extras for provider-neutral consumers.
+        self.extras.setdefault("model", self.model)
+        self.extras.setdefault("batch_endpoint", self.batch_endpoint)
+        self.extras.setdefault("completion_window", self.completion_window)
+        if self.base_url:
+            self.extras.setdefault("base_url", self.base_url)
+        if self.metadata:
+            self.extras.setdefault("metadata", dict(self.metadata))

src/mmirage/core/process/batch/__init__.py

@@ -0,0 +1,28 @@
+"""Provider-agnostic batch processing contracts and registry."""
+
+from mmirage.core.process.batch.adapter import BatchSubmissionAdapter, BatchSubmissionResult
+from mmirage.core.process.batch.collector import collect_and_merge
+from mmirage.core.process.batch.chunking import BatchRequestChunker, RequestChunk
+from mmirage.core.process.batch.openai_adapter import OpenAIBatchAdapter
+from mmirage.core.process.batch.orchestrator import BatchSubmissionOrchestrator
+from mmirage.core.process.batch.registry import BatchAdapterFactory, BatchAdapterRegistry
+from mmirage.core.process.batch.status_checker import (
+    extract_unique_provider_batches,
+    run_status_checker,
+)
+from mmirage.config.openai_batch import OpenAIBatchConfig
+
+__all__ = [
+    "BatchSubmissionAdapter",
+    "BatchSubmissionResult",
+    "collect_and_merge",
+    "BatchRequestChunker",
+    "RequestChunk",
+    "BatchSubmissionOrchestrator",
+    "OpenAIBatchAdapter",
+    "OpenAIBatchConfig",
+    "BatchAdapterFactory",
+    "BatchAdapterRegistry",
+    "extract_unique_provider_batches",
+    "run_status_checker",
+]