Config

Telescope uses a flat, Pydantic-validated configuration schema. Every field name is globally unique and self-descriptive. Typos in YAML keys are caught at load time (extra="forbid").

Config loading

Configuration is resolved by merging three layers:

Defaults — configs/defaults/default_train.yaml (ships with Telescope)
Run config — your YAML file passed via --config
CLI overrides — individual flags like --learning_rate 5e-7

Later layers override earlier ones. You only need to specify the fields you want to change.

uv run train.py --config configs/my_run.yaml --learning_rate 5e-7

Summary

The most commonly configured parameters for a training run:

# Model
model: "Qwen/Qwen2.5-7B"        # HuggingFace model ID or local path
model_dtype: "bfloat16"         # model weight precision

# Environments
environments:
  - name: "hendrycks_math"      # registered environment name
    weight: 1.0                 # sampling weight for multi-env training
    reward_min: 0.0             # expected min reward (for UI charts)
    reward_max: 1.0             # expected max reward (for UI charts)

# Workers
inference_num_workers: 4        # number of vLLM inference servers
trainer_num_workers: 4          # number of FSDP/Megatron trainer workers

# Training
algorithm: "grpo"               # RL algorithm (grpo, rloo, reinforce_pp, dr_grpo, cispo, gspo, sapo)
group_size: 8                   # completions sampled per prompt
learning_rate: 1.0e-6           # optimizer learning rate
prompts_batch_size_for_trainer: 16      # number of groups per training batch
number_of_steps: 300            # total training steps

# Sampling
temperature: 1.0                # sampling temperature
max_tokens: 3700                # max generation tokens per completion
seq_len: 4096                   # max sequence length for packing

# Async
max_async_rollout: 2            # how many steps inference can run ahead (0 = synchronous)

# Checkpointing
checkpoint_every: 50            # save a checkpoint every N steps

# Logging
wandb_project: "my-project"    # Weights & Biases project name
wandb_run_name: "my-run"       # Weights & Biases run name

For the full list of parameters, see the sections below.

General

Parameter	Type	Default	Description
`debug`	`bool`	`false`	Enable debug mode

Model

Parameter	Type	Default	Description
`model`	`str`	`"Qwen/Qwen2.5-3B"`	HuggingFace model identifier or local path
`model_dtype`	`"float32" \| "float16" \| "bfloat16"`	`"float32"`	Model weight precision
`mixed_precision_dtype`	`"float32" \| "float16" \| "bfloat16" \| null`	`"bfloat16"`	FSDP mixed precision compute dtype. `null` to disable

Environments

At least one environment must be configured. Each entry is an object with these fields:

Parameter	Type	Default	Description
`name`	`str`	required	Environment name (must match a registered environment)
`weight`	`float`	`1.0`	Sampling weight for multi-environment training (must be > 0)
`reward_min`	`float \| null`	`null`	Expected minimum reward (used by the UI for normalized charts)
`reward_max`	`float \| null`	`null`	Expected maximum reward (used by the UI for normalized charts)
`kwargs`	`dict`	`{}`	Environment-specific keyword arguments

environments:
  - name: "hendrycks_math"
    weight: 0.5
    reward_min: 0.0
    reward_max: 1.0
  - name: "countdown"
    weight: 0.5
    reward_min: 0.0
    reward_max: 2.0

Workers

Parameter	Type	Default	Description
`inference_num_workers`	`int`	`2`	Number of vLLM inference workers. Must be divisible by `inference_tensor_parallel_size`
`inference_tensor_parallel_size`	`int`	`1`	Tensor parallelism degree for each inference server
`trainer_num_workers`	`int`	`2`	Number of FSDP/Megatron trainer workers

Orchestrator

Parameter	Type	Default	Description
`max_concurrent_prompts_per_server`	`int`	`32`	Max concurrent prompts per inference server
`prompts_batch_size_for_trainer`	`int`	`16`	Number of prompt groups per training batch
`number_of_steps`	`int`	`300`	Total training steps
`max_async_rollout`	`int`	`2`	How many training steps inference can run ahead. `0` = fully synchronous
`discard_group_zero_advantage`	`bool`	`true`	Drop groups where all samples got identical rewards (no learning signal)
`enable_prompt_prefetch`	`bool`	`true`	Prefetch prompts from the dataset
`prompt_prefetch_buffer_size`	`int`	`24`	Number of prompts to prefetch
`enable_individual_sample_lanes`	`bool`	`true`	Each sample gets its own lane slot instead of one per group
`free_lane_after_generation`	`bool`	`true`	Free inference lane after generation completes (before reward computation). Improves throughput when reward computation is slow (e.g., sandbox environments). Requires `enable_individual_sample_lanes`
`max_off_policy_steps`	`int`	`8`	Cancel in-flight rollouts after this many weight updates

Trainer

Parameter	Type	Default	Description
`learning_rate`	`float`	`1.0e-6`	Learning rate (must be > 0)
`weight_decay`	`float`	`0.0`	Weight decay coefficient
`grad_clip`	`float`	`1.0`	Maximum gradient norm for clipping
`lr_scheduler`	`"none" \| "constant" \| "linear" \| "cosine"`	`"constant"`	LR schedule. `"none"` = fixed LR; `"constant"` = warmup then constant; `"linear"` / `"cosine"` = warmup then decay
`warmup_steps`	`int`	`0`	Number of linear warmup steps
`min_lr_ratio`	`float`	`0.0`	Minimum LR as a fraction of `learning_rate` for linear/cosine decay. Range: [0, 1]
`train_backend`	`"fsdp" \| "megatron"`	`"fsdp"`	Training backend

Algorithm

Parameter	Type	Default	Description
`algorithm`	`"grpo" \| "rloo" \| "reinforce_pp" \| "dr_grpo" \| "cispo" \| "gspo" \| "sapo"`	`"grpo"`	RL algorithm. See Algorithms
`number_of_minibatches`	`int`	`1`	PPO-style gradient steps per rollout batch. >1 enables multiple update passes
`advantage_norm`	`"group" \| "batch"`	`"group"`	Normalize advantages within each prompt group or across the training batch
`use_ppo_clip`	`bool`	`false`	Apply PPO ratio clipping. Incompatible with `cispo`, `gspo`, `sapo`
`ppo_clip_ref_logprobs`	`"rollout" \| "batch"`	`"rollout"`	Reference logprobs for PPO clipping. `"rollout"` = vLLM logprobs; `"batch"` = recompute with trainer
`clip_low`	`float`	`0.4`	PPO clip lower bound: ratio is clamped to `1 - clip_low`. Range: [0, 1]
`clip_high`	`float`	`0.5`	PPO clip upper bound: ratio is clamped to `1 + clip_high`
`sapo_tau_pos`	`float`	`1.0`	SAPO sigmoid sharpness for positive advantages
`sapo_tau_neg`	`float`	`1.05`	SAPO sigmoid sharpness for negative advantages
`dr_grpo_loss_agg_mode`	`"token_mean" \| "token_sum_norm"`	`"token_mean"`	DR-GRPO loss aggregation. `"token_sum_norm"` removes response-level length bias
`use_tis`	`bool`	`false`	Truncated importance sampling for async vLLM/trainer logprob correction
`tis_cap`	`float`	`2.0`	Max importance weight for TIS
`tis_logprob_clamp`	`float`	`20.0`	Logprob clamping threshold for TIS
`entropy_chunk_size`	`int`	`1024`	Chunk size for entropy calculation

use_tis=true + use_ppo_clip=true + ppo_clip_ref_logprobs="rollout" is invalid — it double-counts the importance sampling correction. Use ppo_clip_ref_logprobs: "batch" instead.

Rollout / Sampling

Parameter	Type	Default	Description
`group_size`	`int`	`8`	Number of completions sampled per prompt
`temperature`	`float`	`1.0`	Sampling temperature
`top_p`	`float \| null`	`null`	Nucleus sampling probability. `null` = disabled
`max_tokens`	`int`	`3700`	Maximum generation tokens per completion
`interleaved_rollouts`	`bool`	`true`	Exact token reuse across turns — avoids re-tokenization mismatches in multi-turn environments

Sequence Packing

Parameter	Type	Default	Description
`seq_len`	`int`	`4096`	Maximum sequence length for packing
`pad_to_multiple_of`	`int`	`64`	Pad packed sequences to this alignment

Inference Server

Parameter	Type	Default	Description
`inference_host`	`str`	`"0.0.0.0"`	Host address for inference servers
`inference_base_port`	`int`	`8100`	Base port for inference servers
`gpu_memory_utilization`	`float`	`0.9`	Fraction of GPU memory allocated to vLLM. Range: (0, 1]
`max_model_len`	`int`	`4096`	Maximum context length for vLLM
`vllm_scheduling_policy`	`"priority" \| "fcfs"`	`"priority"`	`"priority"` = turn-aware scheduling for multi-turn; `"fcfs"` = first-come-first-served
`enable_thinking`	`bool`	`false`	Pass `enable_thinking=True` to `apply_chat_template()` for models with native thinking support
`chat_template`	`str \| null`	`null`	Jinja2 chat template string to override the tokenizer’s built-in template. Required for base models that don’t include a chat template in their tokenizer

Checkpointing

See Checkpointing for detailed usage.

Parameter	Type	Default	Description
`checkpoint_every`	`int \| bool`	`false`	Save every N steps. `0` or `false` to disable
`checkpoint_save_training_state`	`bool`	`true`	Include optimizer/scheduler state for resume. `false` = weights-only
`resume_from_checkpoint`	`bool \| int`	`false`	`true` = resume from latest; integer = resume from specific step
`checkpoint_dir`	`str \| null`	`null`	Custom checkpoint path. Default: `RUN_DIR/checkpoints`
`checkpoint_keep_last`	`int \| null`	`null`	Keep only the N most recent checkpoints
`checkpoint_keep_every`	`int \| null`	`null`	Always keep checkpoints at these step multiples

Evals

See Evals for detailed usage.

Parameter	Type	Default	Description
`eval_before_training`	`bool`	`true`	Run evaluations before training starts
`eval_after_training`	`bool`	`true`	Run evaluations after training ends
`eval_num_servers`	`int`	`1`	Number of dedicated eval servers reserved from the inference pool. `0` = no periodic evals during training
`eval_start_end_use_all_servers`	`bool`	`true`	Use all inference servers for start/end evals

Each eval entry supports:

Parameter	Type	Default	Description
`name`	`str`	required	Eval name (environment name or dedicated eval like `math500`)
`eval_every`	`int`	`10`	Run this eval every N training steps
`num_samples`	`int`	`-1`	Number of samples to evaluate. `-1` = full dataset
`separate_eval_samples`	`bool`	`false`	Use a separate set of eval samples
`kwargs`	`dict`	`{}`	Environment-specific keyword arguments
`pass_k`	`dict`	`{}`	pass@k / pass^k configuration
`temperature`	`float \| null`	`null`	Override sampling temperature (`null` = inherit from training config)
`top_p`	`float \| null`	`null`	Override nucleus sampling (`null` = inherit)
`max_tokens`	`int \| null`	`null`	Override max tokens (`null` = inherit)

evals:
  - name: "math500"
    eval_every: 10
    num_samples: 500
    kwargs:
      dataset_name: "HuggingFaceH4/MATH-500"
      dataset_split: "test"
    temperature: 1.0
    max_tokens: 3000

Logging

Parameter	Type	Default	Description
`use_wandb`	`bool`	`true`	Enable Weights & Biases logging
`wandb_project`	`str`	`"telescope"`	W&B project name
`wandb_run_name`	`str`	`"telescope_run"`	W&B run name
`wandb_tags`	`list[str]`	`["telescope"]`	W&B tags
`wandb_upload_code`	`bool`	`true`	Upload source code to W&B
`wandb_upload_logs`	`bool`	`true`	Upload trainer/inference logs to W&B
`wandb_upload_logs_detailed`	`bool`	`false`	Include detailed logs in W&B upload
`wandb_upload_logs_stdout`	`bool`	`false`	Include stdout logs in W&B upload
`wandb_code_max_file_size_mb`	`float`	`2.0`	Max file size for code upload
`wandb_code_exclude_patterns`	`list[str]`	`[".git", ".venv", ...]`	Glob patterns to exclude from code upload
`system_metrics_collection_interval_seconds`	`float`	`1.0`	System metrics collection interval
`torch_memory_sample_interval_seconds`	`float`	`0.1`	PyTorch memory sampling interval
`event_tail_window_seconds`	`int`	`60`	Tail window for event aggregation
`event_block_duration_seconds`	`int`	`1800`	Duration of event blocks
`event_upload_interval_seconds`	`int`	`5`	Event upload interval
`metrics_logger_interval_seconds`	`float`	`2.0`	Metrics logger flush interval
`ray_torch_memory_drain_interval_seconds`	`float`	`0.5`	Interval for draining Ray torch memory metrics
`rollout_block_size`	`int`	`500`	Block size for rollout event grouping
`track_gpu_events`	`bool`	`true`	Track GPU utilization events

Ray Cluster

Parameter	Type	Default	Description
`ray_address`	`str`	`"auto"`	Ray cluster address
`ray_auto_start_local`	`bool`	`true`	Auto-start a local Ray cluster if none is running
`ray_namespace`	`str`	`"telescope"`	Ray namespace
`ray_log_to_driver`	`bool`	`true`	Forward worker logs to the driver process
`ray_runtime_env`	`dict \| null`	`null`	Custom Ray runtime environment
`ray_disable_runtime_env_hook`	`bool`	`true`	Disable Ray’s runtime environment hook
`ray_pin_py_executable`	`bool`	`true`	Pin the Python executable path in workers
`ray_propagate_active_venv`	`bool`	`true`	Propagate the active virtualenv to workers
`ray_propagate_run_dir`	`bool`	`true`	Propagate the run directory to workers
`ray_broadcast_init_timeout_s`	`int`	`300`	Timeout for weight broadcast initialization
`ray_broadcast_prefer_loopback_if_single_node`	`bool`	`true`	Prefer loopback interface for single-node broadcasts
`ray_shutdown_on_exit`	`bool`	`false`	Shut down the Ray cluster on exit
`ray_inference_cpus_per_worker`	`float`	`4.0`	CPU cores allocated per inference worker
`ray_trainer_cpus_per_worker`	`float`	`4.0`	CPU cores allocated per trainer worker
`ray_inference_placement_strategy`	`"PACK" \| "SPREAD" \| "STRICT_PACK" \| "STRICT_SPREAD"`	`"PACK"`	Placement strategy for inference workers
`ray_trainer_placement_strategy`	`"PACK" \| "SPREAD" \| "STRICT_PACK" \| "STRICT_SPREAD"`	`"PACK"`	Placement strategy for trainer workers
`ray_placement_timeout_s`	`int`	`900`	Timeout for Ray placement group creation

Weight Broadcasting

Parameter	Type	Default	Description
`weight_broadcast_mode`	`"flattened_bucket" \| "per_tensor"`	`"flattened_bucket"`	Strategy for broadcasting updated weights to inference servers
`weight_broadcast_bucket_mb`	`int`	`256`	Bucket size in MB for flattened broadcast
`weight_broadcast_cpu_staging`	`bool`	`false`	Use CPU staging buffer for weight transfers
`weight_broadcast_pin_memory`	`bool`	`true`	Use pinned CPU memory for faster H2D transfers
`weight_broadcast_free_grad_buffers`	`bool`	`true`	Free Megatron gradient buffers during broadcast (saves ~14 GB)

Megatron

These settings are ignored when train_backend: "fsdp". Only relevant when train_backend: "megatron".

Parameter	Type	Default	Description
`megatron_tensor_parallel_size`	`int`	`1`	Tensor parallelism degree
`megatron_pipeline_parallel_size`	`int`	`1`	Pipeline parallelism degree
`megatron_context_parallel_size`	`int`	`1`	Context parallelism degree
`megatron_expert_parallel_size`	`int`	`1`	Expert parallelism degree (MoE models)
`megatron_global_batch_size`	`int \| null`	`null`	Global batch size. Defaults to `prompts_batch_size_for_trainer`
`megatron_disable_unified_memory_jit`	`bool`	`true`	Disable unified memory JIT
`megatron_optimizer_cpu_offload`	`bool`	`false`	Offload optimizer states to CPU
`megatron_optimizer_offload_fraction`	`float`	`1.0`	Fraction of optimizer state to offload. Range: [0, 1]
`megatron_overlap_cpu_optimizer_d2h_h2d`	`bool`	`true`	Overlap CPU optimizer D2H/H2D transfers
`megatron_use_precision_aware_optimizer`	`bool`	`false`	Use precision-aware optimizer
`megatron_main_grads_dtype`	`str`	`"float32"`	Data type for main gradients
`megatron_main_params_dtype`	`str`	`"float32"`	Data type for main parameters
`megatron_exp_avg_dtype`	`str`	`"float32"`	Data type for Adam exponential average
`megatron_exp_avg_sq_dtype`	`str`	`"float32"`	Data type for Adam exponential average squared
`megatron_grad_reduce_in_fp32`	`bool`	`true`	Reduce gradients in FP32. `false` halves grad buffer memory (~14 GB savings)
`megatron_gradient_checkpointing`	`bool`	`true`	Recompute activations in backward pass (~15-20 GB memory savings)
`megatron_sequence_parallel`	`bool`	`true`	Shard sequence dimension in LayerNorm/dropout (requires TP > 1)
`megatron_use_distributed_optimizer`	`bool`	`true`	Shard optimizer states across DP ranks (requires DP > 1)
`megatron_overlap_grad_reduce`	`bool`	`false`	Overlap gradient allreduce with backward compute
`megatron_use_transformer_engine`	`bool`	`false`	Use TransformerEngine layer spec (fused attention, FP8-ready)
`megatron_fp8`	`bool`	`false`	Enable FP8 compute (requires TransformerEngine + Hopper GPU)

vLLM Tracing

Parameter	Type	Default	Description
`enable_vllm_tracing`	`bool`	`true`	Enable OpenTelemetry tracing for vLLM
`otlp_receiver_port`	`int`	`4318`	OTLP HTTP receiver port. Range: [1, 65535]

Getting Started

Core Concepts

Operations

Config loading

Summary

General

Model

Environments

Workers

Orchestrator

Trainer

Algorithm

Rollout / Sampling

Sequence Packing

Inference Server

Checkpointing

Evals

Logging

Ray Cluster

Weight Broadcasting

Megatron

vLLM Tracing

Getting Started

Core Concepts

Operations

​Config loading

​Summary

​General

​Model

​Environments

​Workers

​Orchestrator

​Trainer

​Algorithm

​Rollout / Sampling

​Sequence Packing

​Inference Server

​Checkpointing

​Evals

​Logging

​Ray Cluster

​Weight Broadcasting

​Megatron

​vLLM Tracing

Config loading

Summary

General

Model

Environments

Workers

Orchestrator

Trainer

Algorithm

Rollout / Sampling

Sequence Packing

Inference Server

Checkpointing

Evals

Logging

Ray Cluster

Weight Broadcasting

Megatron

vLLM Tracing