Unified training recipe catalog

Every draft model JSON under configs/ has at least one typed YAML recipe in this catalog. Choose a recipe by answering three separate questions:

1. When are target features produced? online captures them during the run; offline reads feature files prepared earlier.
2. How are feature supply and training deployed? colocated maps to deployment.mode: local_colocated and lets the trainer read offline files directly; disaggregated uses separate producer and consumer roles.
3. Who owns the online infrastructure? external means the user or scheduler owns Mooncake and SGLang; managed-local means SpecForge starts and stops those services on the local host.

Those decisions map directly to the directory tree:

examples/configs/
├── offline/
│   ├── colocated/
│   └── disaggregated/
└── online/
    ├── colocated/
    └── disaggregated/
        ├── external/
        └── managed-local/

Directory	Feature source	SpecForge roles	Infrastructure ownership
offline/colocated	Precomputed .ckpt files	Trainer only	Not applicable
offline/disaggregated	Precomputed files ingested into shared_dir or Mooncake	Producer + consumer	Storage is configured separately; there is no external/managed-local subdivision
online/colocated	—	—	Reserved; the current runtime does not support online colocated training
online/disaggregated/external	Live SGLang capture	Producer + consumer	User or scheduler starts Mooncake and SGLang
online/disaggregated/managed-local	Live SGLang capture	Producer + consumer	SpecForge starts local Mooncake and SGLang from managed_local

The directory is the source of truth for mode, topology, and service ownership. Some filenames retain historical -online, -offline, or -disaggregated labels so existing recipe identities and run names remain recognizable.

Run any recipe through the one public training entry:

specforge train --config examples/configs/online/disaggregated/external/qwen3-8b-eagle3-disaggregated.yaml

model.draft_model_config may name a local JSON file, a local model directory, or a Hugging Face repository. Fresh EAGLE3, P-EAGLE, and DFlash runs may omit it and derive the draft architecture from the target; see the training guide for layer/block overrides and the distinction between weights-only model.draft_checkpoint_path and full training.resume_from.

Every recipe records its audited process count under deployment.trainer. Multi-process configs self-launch through torch distributed:

specforge train -c examples/configs/online/disaggregated/external/qwen3-30b-a3b-eagle3-online.yaml

The Qwen3-30B-A3B EAGLE3.1 variant uses the same unified entry point:

specforge train -c examples/configs/online/disaggregated/external/qwen3-30b-a3b-eagle3.1-online.yaml

Its draft config enables per-layer RMS normalization before the three captured target hidden states are concatenated and projected. It remains registered as the eagle3 strategy; EAGLE3.1 is a draft-model configuration variant, not a second runtime or launch path.

Recipes under online/disaggregated/managed-local are opt-in, single-node full-stack examples. Their typed managed_local blocks own Mooncake, one or more patched SGLang capture servers, and the trainer GPU allocation; the same specforge train -c ... command starts and cleans up the complete stack. Recipes under online/disaggregated/external omit managed_local, so Mooncake and SGLang remain owned by the user, scheduler, or service platform.

The kimi-k3-dspark-disaggregated.yaml recipe is the external two-node migration of the 64K Kimi K3 continual run. Its dedicated runbook pins the K3 SGLang revision and patch target, preserves the old effective global batch and prompt order, and documents the TP8 capture plus four-rank trainer topology.

deepseek-v4-flash-dspark-disaggregated.yaml trains a DSpark drafter for DeepSeek-V4-Flash-0731 from scratch on ShareGPT, with external capture servers. Its runbook covers the v0.5.18 SGLang capture patch and the bundled deepseek-v4 chat template (the checkpoint ships no Jinja template).

qwen3.8-27b-dflash2-disaggregated.yaml (external services, two nodes) and its managed-local sibling qwen3.8-27b-dflash2-4server-dp4-disaggregated.yaml (one node, four capture servers plus a DP4 trainer) train the DFlash2 drafter in configs/qwen3.8-27b-dflash2.json for Qwen3.8-27B. Their runbook records the server/trainer splits, Mooncake lease, in-flight watermarks and throughput measured on B300 and H200 nodes, and the two-node launcher that starts eight capture servers.

Before running a recipe, update model/data paths and create any referenced offline feature or vocabulary-mapping artifacts. Managed-local recipes intentionally record their GPU allocation and loopback services. External deployments may record stable endpoints in YAML or override them through the environment; credentials, tokens, and node-local identity should not be checked into a portable recipe.

Writing a recipe

The Pydantic models in specforge/config/schema.py are the authoritative schema. This section explains what belongs in each YAML section and records the defaults that matter when writing a recipe. Unknown or misspelled fields are errors; YAML files and dotted CLI overrides go through the same validation.

New checked-in recipes should explicitly set training.strategy, deployment.mode, deployment.trainer.nnodes, deployment.trainer.nproc_per_node, run_id, and output_dir, even when the schema has the same default. A minimal online recipe with externally managed services looks like:

model:
  target_model_path: Qwen/Qwen3-8B
  draft_model_config: configs/qwen3-8b-eagle3.json
  target_backend: sglang
  vocab_mapping_path: cache/vocab_mapping/qwen3-8b.pt

data:
  train_data_path: ./cache/dataset/sharegpt_train.jsonl
  max_length: 4096
  chat_template: qwen

training:
  strategy: eagle3
  max_steps: 10000
  batch_size: 1
  learning_rate: 0.0001

run_id: qwen3-8b-eagle3-disaggregated
output_dir: outputs/qwen3-8b-eagle3-disaggregated

deployment:
  mode: disaggregated
  trainer:
    nnodes: 1
    nproc_per_node: 1
  disaggregated:
    control_dir: outputs/qwen3-8b-eagle3-disaggregated/control
    consumer_state_dir: outputs/qwen3-8b-eagle3-disaggregated/consumer-state
    backend: mooncake
    server_urls:
      - http://127.0.0.1:30000
    mooncake_metadata_server: http://127.0.0.1:35880/metadata
    mooncake_master_server_addr: 127.0.0.1:35551

Paths are resolved from the current working directory. The checked-in recipes assume the command runs from the repository root.

Choose a starting recipe

Workflow	Canonical starting point
EAGLE3 offline, colocated	offline/colocated/qwen3-8b-eagle3-offline.yaml
DFlash offline, colocated	offline/colocated/qwen3-8b-dflash-offline.yaml
Domino offline, colocated	offline/colocated/qwen3-8b-domino-offline.yaml
DSpark offline, colocated	offline/colocated/qwen3-4b-dspark-offline.yaml
EAGLE3 offline, disaggregated	offline/disaggregated/qwen3-8b-eagle3-offline-disaggregated.yaml
EAGLE3 online, external services	online/disaggregated/external/qwen3-8b-eagle3-disaggregated.yaml
DFlash online, managed-local stack	online/disaggregated/managed-local/qwen3-8b-dflash-1server-dp7-disaggregated.yaml
DFlash 2 online, managed-local stack	online/disaggregated/managed-local/qwen3.8-27b-dflash2-4server-dp4-disaggregated.yaml
DFlash 2 online, external services on two nodes	online/disaggregated/external/qwen3.8-27b-dflash2-disaggregated.yaml
Domino online, managed-local stack	online/disaggregated/managed-local/qwen3-8b-domino-multiserver-disaggregated.yaml

The runtime derives online/offline mode from the selected data source and reads topology from deployment.mode; it does not parse the filename or its historical suffixes.

Top-level fields

Field	Default	What to write
run_id	specforge-run	Stable identifier for the run. It names checkpoints and is also the default disaggregated store namespace. Use a new value for a fresh attempt.
output_dir	./output	Shared checkpoint, profiler, and tracker output directory. Every trainer rank must resolve it to the same location.

The top-level model and data sections are required. training, tracking, profiling, runtime, and deployment have defaults, but checked-in recipes should make their training strategy and topology explicit.

model: target, draft, and capture backend

Field	Default	What to write
model.target_model_path	required	Local target directory or Hugging Face repository ID.
model.draft_model_config	null	Draft JSON, model directory containing config.json, or Hugging Face repository. EAGLE3, P-EAGLE, and DFlash may omit it and derive a fresh config; Domino and DSpark require one.
model.draft_checkpoint_path	null	Weights-only warm start for a new run. Do not combine it with training.resume_from.
model.draft_num_hidden_layers	null	Positive fresh-architecture override where the strategy permits it. EAGLE3 remains one layer; P-EAGLE and DFlash may override their generated defaults.
model.draft_block_size	null	Positive DFlash block-size override; generated DFlash configs default to 16.
model.target_backend	sglang	sglang is the only accepted value; retired hf/custom names fail at config load. Offline feature consumers do not instantiate a target inference backend.
model.input_modality	text	The provider modality. The unified runtime supports text only; VLM modalities such as qwen2_5_vl are rejected.
model.shard_target_output	false	Retained for config migration; leave it false on the online disaggregated path.
model.trust_remote_code	false	Enable only for model repositories that require custom loading code.
model.use_liger_kernel	false	Enable Liger Qwen3 RMSNorm/SwiGLU kernels for DFlash training. Requires the specforge[liger] extra.
model.embedding_key	model.embed_tokens.weight	Target checkpoint key copied into or used by the draft embedding.
model.lm_head_key	lm_head.weight	Target checkpoint key used for the frozen output head.
model.vocab_mapping_path	""	Target-to-draft vocabulary mapping. EAGLE3 disaggregated runs require an explicit shared file.
model.load_target_embedding	true	Copy the frozen target embedding into a fresh draft when supported.
model.aux_hidden_state_layer_ids	null	Optional EAGLE3/P-EAGLE capture override containing exactly three non-negative layer IDs. Other strategies derive layers from the draft config.
model.torch_dtype	bfloat16	bfloat16, float16, or float32.
model.cache_dir	null	Model/tokenizer download cache. This is distinct from data.cache_dir.
model.mask_token_id	null	DFlash-family/P-EAGLE mask token override. Otherwise it resolves from the draft config and then the tokenizer.
model.tokenizer_pad_token_id	null	Explicit non-negative tokenizer pad ID. Use it for released tokenizers that omit padding metadata.
model.sglang_attention_backend	flashinfer	SGLang attention implementation for an in-process or managed capture server.
model.sglang_mem_fraction_static	0.4	SGLang static-memory fraction in (0, 1]; inherited by managed capture servers unless they override it.
model.sglang_disable_radix_cache	true	Preserve the historical managed-capture behavior. Set false for hybrid targets such as Inkling that require the radix tree. Unique per-attempt cache namespaces still force complete capture prefills.
model.sglang_context_length	null	Positive explicit context limit. Managed capture requires at least data.max_length + 7; omitting it derives that value.
model.sglang_enable_nccl_nvls	false	Pass the matching SGLang NCCL NVLS optimization flag.
model.sglang_enable_symm_mem	false	Pass the matching SGLang symmetric-memory flag.
model.sglang_enable_torch_compile	false	Enable the SGLang torch-compile path.
model.sglang_enable_dp_attention	false	Enable SGLang DP attention where supported. Managed-local capture currently rejects it.
model.sglang_enable_dp_lm_head	false	Enable SGLang DP LM head where supported. Managed-local capture currently rejects it.
model.sglang_ep_size	1	SGLang expert-parallel size; it must divide and not exceed every managed capture server's tp_size.
model.sglang_max_running_requests	null	Positive SGLang request-concurrency limit.
model.sglang_max_total_tokens	null	Positive SGLang token-pool limit.
model.sglang_dp_size	null	Optional SGLang data-parallel size.
model.sglang_moe_a2a_backend	null	Optional SGLang MoE all-to-all backend name.
model.sglang_moe_runner_backend	null	Optional SGLang MoE runner backend name.
model.sglang_page_size	null	Optional positive SGLang KV-cache page size.
model.sglang_quantization	null	Optional SGLang target quantization mode.
model.sglang_fp4_gemm_runner_backend	null	Optional SGLang FP4 GEMM runner backend.
model.sglang_mamba_radix_cache_strategy	null	Optional hybrid Mamba/radix cache strategy.
model.sglang_max_mamba_cache_size	null	Optional positive Mamba cache size.
model.sglang_swa_full_tokens_ratio	null	Optional SGLang sliding-window full-token ratio in (0, 1].
model.sglang_mamba_full_memory_ratio	null	Optional SGLang Mamba full-memory ratio in (0, 1].

data: choose exactly one training source

Exactly one of the first three fields must be non-empty:

Field	Default	What to write
data.train_data_path	""	Raw conversation/preformatted JSON or JSONL sent to the online capture producer.
data.prompts_path	""	Pre-tokenized online JSONL with input_ids and loss_mask.
data.hidden_states_path	""	Directory of precomputed offline feature .ckpt files. Selecting it makes the run offline.
data.eval_data_path	""	Reserved migration field. Online evaluation is unsupported; leave it empty.
data.eval_hidden_states_path	""	Offline evaluation features; configure them together with a positive training.eval_interval.
data.max_length	2048	Maximum token length used by preparation, capture, and training.
data.chat_template	llama3	Template name used to format conversations and locate assistant loss spans.
data.is_preformatted	false	Treat each record's text as already formatted by chat_template.
data.train_only_last_turn	false	Restrict the loss mask to the final assistant turn.
data.build_dataset_num_proc	8	Positive CPU process count for dataset preprocessing.
data.dataloader_num_workers	null	Ordered feature-loader workers. null preserves strategy defaults: EAGLE/P-EAGLE 4, DFlash-family 8; use 0 for synchronous loading.
data.cache_dir	./cache	Prepared dataset and derived vocabulary-mapping cache.
data.cache_key	null	Optional explicit namespace when multiple preparations share the same source.
data.max_prompts	null	Optional non-negative prompt cap, useful for smoke tests.

Offline evaluation uses eval_hidden_states_path; configure it together with training.eval_interval. Online evaluation is unsupported, and setting eval_data_path fails config validation.

training: optimization, strategy, and parallelism

Common fields:

Field	Default	What to write
training.strategy	eagle3	eagle3, peagle, dflash, domino, or dspark.
training.num_epochs	1	Positive passes over a finite source.
training.max_steps	null	Positive hard stop in optimizer steps. If it is set while total_steps is omitted, it is also the fallback schedule horizon.
training.total_steps	null	Positive optimizer/loss schedule horizon; it does not itself stop an online stream. A finite online disaggregated run may omit both fields: the producer publishes the exact horizon derived from prepared prompts, epochs, DP size, batch size, and accumulation.
training.batch_size	1	Per-rank microbatch size. P-EAGLE and USP require 1.
training.accumulation_steps	1	Positive microbatches per optimizer update.
training.fsdp_sharding	SHARD_GRAD_OP	Trainer FSDP mode: SHARD_GRAD_OP, FULL_SHARD, or NO_SHARD.
training.learning_rate	1e-4	Positive peak learning rate.
training.lr_scheduler	cosine	Learning-rate schedule after warmup: cosine or constant.
training.warmup_ratio	0.015	Fraction in [0, 1] used for scheduler warmup.
training.max_grad_norm	0.5	Positive gradient-clipping norm.
training.optimizer_cpu_offload	false	Keep the optimizer's FP32 master parameters and Adam state on CPU.
training.attention_backend	flex_attention	eager, sdpa, flex_attention, fa, or usp; the selected strategy must support it.
training.tp_size	1	Online disaggregated consumers must keep it at 1; configure target TP on capture servers. Offline non-USP ranks consume disjoint data.
training.sp_ulysses_size	1	Ulysses sequence-parallel factor for offline EAGLE3 USP.
training.sp_ring_size	1	Ring sequence-parallel factor for offline EAGLE3 USP.
training.dist_timeout	10	Positive distributed-operation timeout in minutes.
training.save_interval	0	Save every N optimizer steps; 0 disables periodic saves. A final checkpoint is still written.
training.eval_interval	0	Evaluate every N optimizer steps; 0 disables evaluation.
training.log_interval	50	Positive optimizer-step logging interval.
training.max_checkpoints	0	Keep the newest N checkpoints; 0 keeps all.
training.resume_from	null	Full-run checkpoint/run root: draft, optimizer/scheduler, counters, data position, and RNG. Mutually exclusive with model.draft_checkpoint_path.
training.compact_teacher	false	Exact lower-peak-memory teacher projection for offline text EAGLE3.
training.compact_teacher_chunk_size	null	Positive vocabulary chunk size; requires compact_teacher: true.
training.trim_loss_positions	false	EAGLE3 only. Compute the teacher target_p, draft logits, and loss only at supervised positions (batch size 1, plain KL loss); mathematically equivalent to the full-length path.
training.role	all	Use all for offline colocated training; disaggregated entrypoints select auto, producer, or consumer.
training.seed	42	Run and per-rank RNG seed.
training.prompt_seed	null	Optional online prompt-shuffle seed. null preserves the historical behavior of using training.seed.

Strategy-specific fields should be written only when tuning that objective:

Strategy	Fields and defaults
EAGLE3	training.ttt_length (7), training.lk_loss_type (null; lambda, alpha, or tv), training.kl_scale (1.0), training.kl_decay (1.0)
DFlash / DFlash 2 / Domino / D-PACE	training.num_anchors (512), training.loss_decay_gamma (null), training.objective_chunk_blocks (128; 0 materializes all objective logits), training.loss_type (dflash; fixed decay, or dpace; dynamic weighting), DFlash/DFlash 2's training.lk_loss_type (null; CE, lambda, alpha, or tv), training.kl_scale (1.0), training.kl_decay (1.0), DFlash 2's CE selector objective controls training.dflash2_selector_loss_alpha (1.0), training.dflash2_selector_warmup_ratio (0.0), training.dflash2_selector_ramp_ratio (0.0), and training.dflash2_selector_stop_gradient (false), training.dpace_alpha (0.5), training.lambda_base_start (1.0), training.lambda_base_decay_ratio (0.5)
DSpark	Token-pooled objective with valid-first-target anchors and distributed ratio telemetry. Configure the shared training.num_anchors (512), training.loss_decay_gamma (null; production recipes use 4.0), and training.objective_chunk_blocks (128; 0 materializes all objective logits), plus training.dspark_ce_loss_alpha (0.1), training.dspark_l1_loss_alpha (0.9), and training.dspark_confidence_head_alpha (1.0).
P-EAGLE	training.num_depths (8), training.down_sample_ratio (0.8), training.down_sample_ratio_min (0.2), training.norm_before_residual (null)
MTP	training.mtp_objective_chunk_size (4096 token positions per lm_head + cross-entropy chunk; 0 disables chunking).

New recipes must not write the loader-only migration fields training.deployment_mode, training.server_urls, or training.metadata_db_path; use the typed deployment.* surface below.

deployment: process and service topology

Field	Default	What to write
deployment.mode	local_colocated	Set local_colocated or disaggregated explicitly in every new recipe.
deployment.trainer	default object	Trainer process topology described by the fields below.
deployment.disaggregated	null	Required object only when deployment.mode: disaggregated.
deployment.trainer.nnodes	1	Number of trainer nodes.
deployment.trainer.nproc_per_node	1	Trainer processes per node; the single CLI self-launches local ranks.
deployment.trainer.node_rank	null	Node-local rank. Shared multi-node YAML normally omits it and passes --node-rank.
deployment.trainer.master_addr	null	Rendezvous address; required when nnodes > 1.
deployment.trainer.master_port	29500	Rendezvous port.

The trainer world size is nnodes * nproc_per_node. training.tp_size must remain 1, and the world size must be divisible by training.sp_ulysses_size * training.sp_ring_size.

For deployment.mode: disaggregated, also write:

Field	Default	What to write
deployment.disaggregated.control_dir	required	Fresh attempt-scoped directory for refs/manifest and lifecycle markers. Shared by default; with inbox_server_url, only producer and consumer rank 0 must share it.
deployment.disaggregated.backend	required	mooncake or shared_dir. Online disaggregated runs require Mooncake.
deployment.disaggregated.consumer_state_dir	null	Node-local rank-0 SQLite/WAL root. Required for multi-node online consumers; their rank inboxes remain under shared control_dir.
deployment.disaggregated.inbox_server_url	null	Optional private http://host:port rank-0 relay for tensor-free inbox refs when remote trainer ranks cannot share control_dir. Online multi-node only; no credentials, path, query, TLS, or built-in authentication.
deployment.disaggregated.store_root	null	Shared feature directory; required when backend: shared_dir.
deployment.disaggregated.store_id	null	Feature-store namespace; defaults to run_id.
deployment.disaggregated.server_urls	[]	External patched SGLang capture endpoints. One rollout worker is created per entry. Do not set with managed_local.
deployment.disaggregated.mooncake_metadata_server	null	External Mooncake metadata URL.
deployment.disaggregated.mooncake_master_server_addr	null	External Mooncake RPC host:port.
deployment.disaggregated.mooncake_local_hostname	null	Node-local Mooncake transfer hostname; usually supplied through the environment.
deployment.disaggregated.mooncake_protocol	null	External transfer protocol such as tcp or rdma.
deployment.disaggregated.mooncake_rdma_devices	null	External Mooncake RDMA-device selection.
deployment.disaggregated.producer_segment_size	null	Positive allocation owned by an offline Mooncake producer. Online capture is server-owned and forces client segments to zero.
deployment.disaggregated.client_buffer_size	268435456	Per-role Mooncake client buffer in bytes.
deployment.disaggregated.idle_timeout_s	null	Positive consumer idle timeout.
deployment.disaggregated.peer_wait_timeout_s	null	Optional positive producer/consumer peer-completion timeout. Unset is unbounded; expiration fails the attempt.
deployment.disaggregated.producer_hold_s	null	Optional positive offline producer retention timeout. Unset is unbounded; expiration fails the attempt.
deployment.disaggregated.shutdown_grace_s	30.0	SIGTERM-to-SIGKILL window for a plain supervisor teardown; must cover worker cleanup (Mooncake drains, checkpoint flush, failure sentinels). managed_local stacks use managed_local.shutdown_grace_s.
deployment.disaggregated.managed_local	null	Optional owned single-node Mooncake + capture-server stack described below.

The four path fields have different ownership:

Path	Lifetime and visibility
output_dir	Shared durable checkpoints and run outputs.
deployment.disaggregated.control_dir	Shared, fresh attempt control state.
deployment.disaggregated.consumer_state_dir	Fresh node-local online-consumer ledger; required for multi-node runs.
deployment.disaggregated.store_root	Shared offline feature payloads for shared_dir.

An external-service online run writes server_urls and Mooncake endpoints in YAML, or injects their environment equivalents. A managed-local run replaces those fields with one owned stack:

deployment:
  mode: disaggregated
  trainer:
    nnodes: 1
    nproc_per_node: 2
  disaggregated:
    control_dir: ./outputs/domino/control
    backend: mooncake
    managed_local:
      trainer_cuda_visible_devices: ["2", "3"]
      mooncake:
        protocol: tcp
      capture_servers:
        - port: 30000
          cuda_visible_devices: ["0"]
          tp_size: 1
        - port: 30001
          cuda_visible_devices: ["1"]
          tp_size: 1

Managed-local fields:

Field	Default	What to write
deployment.disaggregated.managed_local.trainer_cuda_visible_devices	required	One device token per nproc_per_node; trainer and capture devices must not overlap.
deployment.disaggregated.managed_local.mooncake	default object	Owned loopback Mooncake configuration described by the nested fields below.
deployment.disaggregated.managed_local.capture_servers	required	One or more owned patched SGLang server definitions.
deployment.disaggregated.managed_local.shutdown_grace_s	30	Positive graceful process-group shutdown window.
deployment.disaggregated.managed_local.mooncake.rpc_port	35551	Owned Mooncake RPC port.
deployment.disaggregated.managed_local.mooncake.metadata_port	35880	Owned metadata HTTP port.
deployment.disaggregated.managed_local.mooncake.metrics_port	35903	Owned metrics port.
deployment.disaggregated.managed_local.mooncake.local_hostname	127.0.0.1	Local transfer hostname.
deployment.disaggregated.managed_local.mooncake.protocol	tcp	tcp or rdma.
deployment.disaggregated.managed_local.mooncake.rdma_devices	null	RDMA-device selection when using RDMA.
deployment.disaggregated.managed_local.mooncake.global_segment_size_bytes	34359738368	Owned global segment size.
deployment.disaggregated.managed_local.mooncake.local_buffer_size_bytes	1073741824	Owned local client buffer.
deployment.disaggregated.managed_local.mooncake.startup_timeout_s	60	Positive Mooncake readiness timeout.
deployment.disaggregated.managed_local.mooncake.default_kv_lease_ttl_ms	500	Master key-lease TTL (ms) forwarded to mooncake_master --default_kv_lease_ttl. Kept below the consumer's teardown drain window so managed_local shuts down cleanly; set null to inherit Mooncake's stock default.
deployment.disaggregated.managed_local.capture_servers[].port	required	Unique capture HTTP port.
deployment.disaggregated.managed_local.capture_servers[].cuda_visible_devices	required	Device tokens for this server. Their count must equal its tp_size.
deployment.disaggregated.managed_local.capture_servers[].tp_size	1	Target-model tensor parallelism for this server.
deployment.disaggregated.managed_local.capture_servers[].mem_fraction_static	null	Optional SGLang static-memory override in (0, 1]; otherwise inherit model.sglang_mem_fraction_static.
deployment.disaggregated.managed_local.capture_servers[].attention_backend	null	Server-specific override; otherwise inherit model.sglang_attention_backend.
deployment.disaggregated.managed_local.capture_servers[].startup_timeout_s	1800	Positive server readiness timeout.

Managed-local is only for a fresh, single-node, online Mooncake run. It derives server URLs and Mooncake endpoints, so do not combine it with explicit external endpoints, store_root, or producer_segment_size. It does not support resume, an existing torchrun, or --node-rank. All owned ports and GPU assignments must be disjoint.

runtime: streaming backpressure

This section affects disaggregated streaming producers and is normally omitted unless tuning throughput or memory pressure.

Field	Default	What to write
runtime.producer_lease	8	Prompts leased to a rollout worker at once.
runtime.producer_concurrency	1	Concurrent capture calls maintained by each server's logical producer. Increase to keep ingress full without duplicating producers.
runtime.in_flight_high_watermark	256	Pause production at this many committed, unacknowledged refs.
runtime.in_flight_low_watermark	192	Resume production at or below this count; it cannot exceed the high watermark.
runtime.resident_high_watermark_bytes	null	Optional byte-level pause threshold.
runtime.resident_low_watermark_bytes	null	Optional byte-level resume threshold; requires and cannot exceed the resident high watermark.
runtime.feature_store_max_resident_bytes	null	Optional hard store budget; it cannot be smaller than the resident high watermark.
runtime.feature_store_max_quarantined_bytes	8589934592	Per-process Mooncake receive-buffer quarantine limit; exceeding it fails loudly.

tracking: experiment logging

Field	Default	What to write
tracking.report_to	none	none, wandb, tensorboard, swanlab, or mlflow.
tracking.wandb_project	null	W&B project.
tracking.wandb_name	null	W&B run name.
tracking.wandb_key	null	W&B API key; prefer the environment instead of committing it.
tracking.wandb_offline	false	Use W&B offline mode.
tracking.wandb_dir	null	W&B local state directory.
tracking.swanlab_project	null	SwanLab project.
tracking.swanlab_name	null	SwanLab run name.
tracking.swanlab_key	null	SwanLab API key; prefer the environment.
tracking.mlflow_tracking_uri	null	MLflow tracking endpoint.
tracking.mlflow_experiment_name	null	MLflow experiment.
tracking.mlflow_run_name	null	MLflow run name.

profiling: bounded per-rank traces

Field	Default	What to write
profiling.enabled	false	Enable PyTorch profiler traces on trainer ranks. Capture-only producer roles cannot enable it.
profiling.start_step	30	First completed optimizer step to profile.
profiling.num_steps	4	Positive number of optimizer steps to record.
profiling.record_shapes	false	Include tensor-shape metadata at additional overhead.

Cross-section checks

• Offline feature consumers require training.tp_size: 1. Non-USP ranks each consume a disjoint data shard; USP peers share a sequence within their SP group while draft-DP groups remain disjoint.
• Online disaggregated runs require model.target_backend: sglang and backend: mooncake. When both step fields are omitted, the producer publishes the finite prompt plan's optimizer horizon and the consumer trains to EOF; max_steps remains an optional hard cap. Every trainer rank is data parallel, so training.tp_size, training.sp_ulysses_size, and training.sp_ring_size must remain 1; configure target TP on each capture server.
• USP is offline EAGLE3 only, requires training.batch_size: 1, and requires sp_ulysses_size * sp_ring_size > 1. Non-USP runs keep both SP sizes at 1.
• P-EAGLE reuses the EAGLE3 server feature schema, uses flex_attention, and requires batch size 1.
• VLM training, including Qwen2.5-VL, is not supported. Online capture accepts text inputs only.
• training.compact_teacher is offline text EAGLE3 only.
• Online evaluation is not supported. Offline data.eval_hidden_states_path and training.eval_interval must be configured together.

After copying a checked-in recipe to ./my-run.yaml and editing it, validate the complete schema and inspect the resolved processes without starting a run:

specforge train -c ./my-run.yaml --plan

Use validated dotted overrides for temporary changes:

specforge train -c ./my-run.yaml \
  training.learning_rate=5e-5 \
  'deployment.disaggregated.server_urls=["http://capture-0:30000"]'

For deeper lifecycle and recovery semantics, see the training guide and disaggregated training guide.

Capability matrix

Strategy	Online disaggregated	Offline colocated	Offline disaggregated
EAGLE3	consumer DP	DP + USP	consumer DP
DFlash	consumer DP	DP	consumer DP
Domino	consumer DP	DP	consumer DP
DSpark	consumer DP	DP	consumer DP
P-EAGLE	consumer DP, batch size 1	No	No

qwen3-8b-dpace-online.yaml is the D-PACE recipe. It deliberately uses the shared DFlash strategy with training.loss_type: dpace; D-PACE is an objective selection inside the unified trainer, not another training entry.

DFlash 2 treats the token objective and position weighting as independent A/B axes. training.lk_loss_type: null keeps hard-target CE, tv minimizes the one-hot total-variation objective, and lambda adaptively mixes CE with TV; alpha is equivalent to CE for these hard targets. Independently, training.loss_type: dflash uses loss_decay_gamma, while dpace applies one detached dynamic position weight to both the unary and candidate-selector objectives. For example, LK-lambda with D-PACE uses:

training:
  lk_loss_type: lambda
  kl_scale: 1.0
  kl_decay: 1.0
  loss_type: dpace
  dpace_alpha: 0.5

Evaluation is currently offline-only and pairs training.eval_interval with data.eval_hidden_states_path. Best checkpoints are linked as <run_id>-best. Offline text EAGLE3 may enable training.compact_teacher; tracking.report_to selects none, W&B, TensorBoard, SwanLab, or MLflow.

Ascend NPU launch

Install vendor-matched PyTorch, torch_npu, Mooncake, and an NPU-compatible SGLang capture server first. The *-npu.yaml consumers use SDPA while target capture remains outside the trainer. For example:

export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3
export HCCL_CONNECT_TIMEOUT=7200
export HCCL_EXEC_TIMEOUT=7200
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

specforge train -c examples/configs/online/disaggregated/external/qwen3.5-4b-dflash-online-npu.yaml

The unified launcher provides rank/world/rendezvous variables and the runtime selects HCCL when torch_npu is active. For AMD GPUs, install requirements-rocm.txt; HF + SDPA is the portable ROCm starting point.

Offline colocated and offline disaggregated resume are supported. Disaggregated online recovery resumes only the consumer against retained control/data-plane state; capture producers always start a fresh attempt.

Migration notes:

• The former run_qwen3_8b_dflash_disagg_1srv_dp7.sh self-contained topology is retained as qwen3-8b-dflash-1server-dp7-disaggregated.yaml: one managed capture server on GPU 0 and seven DFlash trainer ranks on GPUs 1–7.
• The former run_qwen3_8b_domino_disagg_1srv_dp7.sh self-contained topology is retained as qwen3-8b-domino-1server-dp7-disaggregated.yaml: one managed capture server on GPU 0 and seven trainer ranks on GPUs 1–7.
• The former run_qwen3.6_27b_dflash_disagg.sh one-server topology is retained as qwen3.6-27b-dflash-1server-dp2-disaggregated.yaml: one managed capture server on GPU 0 and two trainer ranks on GPUs 1–2. The external-service YAML remains available for scheduler-managed deployments.
• The latest pre-cleanup run_qwen3_8b_domino_disagg_multiserver.sh had been reduced to one SGLang server and one URL despite its historical name. The managed Qwen3-8B Domino recipe above restores a genuine two-server topology without restoring the legacy trainer script; the external-service Domino recipe also accepts any number of typed deployment.disaggregated.server_urls.
• The former GPT-OSS-120B shell accidentally selected the 20B draft config; gpt-oss-120b-eagle3-online.yaml points to the matching 120B config.
• The former Qwen3-235B shell accidentally launched Qwen3-Next-80B; the two now have separate recipes.
• Qwen3-Next online EAGLE3 retains its batch size of two. P-EAGLE requires batch size one.
• The old Qwen3.5-35B offline shell had its training command commented out. The YAML records that intended offline trainer configuration after feature preparation.
• Qwen3-8B DTA still shares the DFlash trainer, as before. Its specialized behavior is encoded by the draft JSON (training_mode: vp_drafter); there is no second DTA training entry.