Get a patron from the database.

Get the product from the database. Args: product_id: The product id, such as '6086499569'. Be careful the product id is different

Get valid groups including timeout-based groups

Get the variant from the database. Args: product_id: The product id, such as '6086499569'. Be careful the product id is different

Detect whether W&B should run in offline mode. Priority order: 1) args.wandb_mode if provided 2) WANDB_MODE environment variable

Check if the order is pending. This is not a strict check, and not meant to be used for modify_items in pending orders. Args: ord

Non-strict pending check for a sale.

Ported directly from MemAgent hotpotqa.py; supports nested braces and \\boxed space format.

Normalize answer expressions.

Normalize whitespace, punctuation, casing for robust comparison.

Process payment for update reservation Args: user: The user to process payment for. payment_id: The payment

Attempt to get the local LAN IP for the specific family (IPv4/IPv6). Strategy: UDP Probe (Preferred) -> Hostname Resolution (Fallback

Helper method to set up routes for a set of tools

Fully aligned with strip_string in the original MemAgent hotpotqa.py.

Remove <think>...</think> blocks for concise display, keep tool_call and message.

The tools field in the data may be a single dict (func_call) or a list (qa); normalize to a list in either case.

Toggles the Wi-Fi Calling setting. Returns the new state.

Truncate text and add ellipsis if too long.

Gather EP → HF → broadcast. Clears buffer.

Wrap IPv6 address in [] if needed.

Add samples to the data source

All-gather TP-sharded param to full tensor. expert_bias→param, non-TP/duplicated→param.data. Uses expert-TP for ".experts.", else regular-TP.

* Analyze current pathname to figure out repo root + language segment pattern. * Supports patterns: * /en/… (language as first s

Apply FSDP v2 to the model. Args: model: The model to wrap with FSDP mesh: Optional DeviceMesh for FSDP. If None, uses all ranks.

Wrap a given function with additional predefined arguments into a Tool. This function allows converting a standard function into a 'Tool' by

Allocate GPU resourced and initialize model, optimzier, local ckpt, etc.

Build QA eval examples from frames.jsonl.

Call orchestrator via SGLang /generate (matches training path exactly). Training uses tokenizer.apply_chat_template + SGLang /generate + XML pars

Perform ceiling division of two integers. Args: x: the dividend. y: the divisor. Returns: The resul

Helper function to handle domain selection

Compute Generalized Advantage Estimation (GAE) using a FlashLinearAttention- inspired algorithm: parallel prefix scan within chunks and recur

Compute advantages and returns in-place based on `args.advantage_estimator`. This function extracts rewards, log-probs, values, and masks from

Compute GSPO-style per-sequence KL divergence. Args: full_log_probs: Current policy log-probs per sample (full or CP-local). full

Compute Off-Policy Sequence Masking (OPSM) mask. Args: args: Configuration containing `opsm_delta` threshold. full_log_probs: Cur

Create NCCL group: training rank 0 + all engine GPUs. Blocks until joined. ``engine_gpu_counts`` gives the number of GPUs per engine. When

Convert Qwen3 Next model parameters from Megatron to HuggingFace format.

Create placement groups for actor and rollout engines.

Destroy NCCL on training and engines.

Display a single message in markdown format.

Enable forward pre-hooks for provided DDP-wrapped model chunks. Args: model_chunks (Sequence[DDP]): Sequence of DDP modules to enable hoo

Structured mode: send full conversation history, optionally with tools schema for native tool calling (func_call tasks). Arg

Extract the answer from inside a LaTeX \\boxed{} command

Format the time in the format YYYYMMDD_HHMMSS.

Run forward passes only and collect non-loss outputs (e.g., logprobs). The model is put into evaluation mode, a forward-only pipeline pass is

Perform garbage collection based on weight version. Remove nodes with weight_version < (current_weight_version - gc_threshold_k).

Generates a random string of specified length using alphanumeric characters.

Generate a CP-ready micro-batch with packed sequence parameters. Steps: - Fetch raw fields via iterator. - Save original token tenso

Create iterators and a micro-batch schedule for a rollout step. - If `use_dynamic_batch_size` is False, splits into fixed-size contiguous

Get a hash of the agent database Returns None if the database is not available

Get the hash of the database.

Generate a unique hash for dict. Returns a hex string representation of the hash.