Repository navigation
TL;DR
v0.2 at a glance
Runs anywhere
- Four hardware platforms: NVIDIA, AMD, TPU and Trainium (landing soon).
new - torchtitan as the third training backend, next to Megatron-LM and FSDP.
since v0.1.2
Stable on messy fleets
- Top-p mask replay scores on the exact candidate set the sampler drew from.
since v0.1.1 - Score centering removes the drift that stale samples add to the gradient.
since v0.1.2 - Fault tolerance (experimental): a failed cell no longer ends the run.
since v0.1.3
Any workload
- Multi-LoRA: many adapters on one base, Tinker API compatible, now on native adapters.
extended - Multimodal RL in token space: image, video and audio through one media processor, so rollout and trainer see the same tokens.
new - Miles-Diffusion: async training, new models (Krea-2, Qwen-Image 2.1), score centering and DMD2 distillation.
new
Also in this release
Muse Spark Open and MiniMax-M3 join the main image. Muse Spark Open (816B omni MoE) now trains from the stock radixark/miles image with 64×H200 and 32×B300 launchers, replacing the separate muse-spark-open image. MiniMax-M3 and MiniMax-M3.1-Flash are new. DeepSeek-V4.1-Flash gains native LoRA, and GLM-5.3-Flash gains a full-parameter torchtitan path.
Mix datasets, store rollout outputs off the HTTP path. --rollout-data-yaml-config draws every rollout step from several datasets in a fixed ratio, each with its own reward, agent function and length caps, with checkpointed positions. --use-output-store moves routing-replay, indexer and sampling-mask arrays from the /generate response into the run's Mooncake store.
NIXL and a typed configuration. P2P weight transfer gains NIXL as a second transport. Every Miles-owned CLI argument is now a typed configuration field, immutable after parsing outside declared mutable scopes, miles builds as a pure wheel on Python 3.11+, and torchtitan moves to v0.4.0.
Upgrade carefully. --rollout-max-context-len is gone (use --max-seq-len), rollout prompts are shuffled by default, rollout/raw_reward changed meaning, parse_args() returns an immutable AllConfig, --lora-type defaults per backend, Tinker clients must run SDK 0.36.0, torchtitan v0.3.0 checkpoints do not resume, and Python 3.10 is no longer supported. See Breaking Changes & Upgrade Notes.
Highlights
400 PRs from 20 contributors.
Model Support
| Model | Training | What changed | PRs | Docs |
|---|---|---|---|---|
| Muse Spark Open (816B-A42B) | Full (Megatron raw), BF16 | Model code moves into the main image; 64×H200 and 32×B300 launchers; text-only | #4759, #4117 @guapisolo | link (page not yet updated, see below) |
| MiniMax-M3 | Full (Megatron), BF16 | Newly supported; full model on 8×4 GB300 | #4581, #4582 @Zhichenzzz | link |
| MiniMax-M3.1-Flash | Full (Megatron), BF16 from a dequantized copy | Newly supported on the M3 plugin; 4-layer runs only | #4583, #4584 @Zhichenzzz | link |
| Qwen3.5-9B / Qwen3.5-35B-A3B (image, video); Qwen3.8-27B (image) | Full text model, frozen vision tower (Megatron raw) | New mode: multimodal RL in token space; TimeLens video and geo3k recipes on Qwen3.5, a 2-rollout geo3k check on Qwen3.8-27B | #4240, #4272, #4610, #4633, #4742, #4746, #4749 @Rockdu | TimeLens, geo3k |
| Inkling-Small (image, audio) | LoRA (Megatron raw) | New mode: audio RL (AVQA); Inkling multimodal is now token-space only | #4732 @Rockdu | AVQA, Inkling |
| DeepSeek-V4.1-Flash | LoRA (native) | New mode: native LoRA on attention and the shared expert | #4205 @yueming-yuan | link |
| GLM-5.3-Flash | Full (torchtitan) | New backend: FSDP+EP (PP and CP on the 4-layer slice) with CPU offload and NVMe Adam; 313B on 32×H200 | #3989, #4539 @Zhichenzzz, #3995, #4439 @HJSang | link |
| MiMo-V2.6-Flash | Multi-LoRA (Bridge, Tinker gateway) | New mode: concurrent LoRA tenants on a bf16 / fp8 / fp8-mxfp4 rollout base | #4326 @guapisolo | link |
| Qwen3-VL MoE (Bridge) | Full | Fix: the whole model trains again (only the vision mergers trained before) | #4331 @Rockdu | — |
| Nemotron-H / Nemotron 3 (hybrid Mamba) | Full | Fix: Mamba state reset at packed-sample boundaries in THD | #4222 @omri-manolo | — |
| Kimi-K3.1 | W4A16 LoRA | Recipe page only; code and launchers live on radixark/miles:kimi-k31, not in this release |
#4665 @guapisolo | link |
Muse Spark Open in the main image. The 816B omni MoE's 62-layer language model trains natively in Megatron from the stock image: miles_plugins/models/muse_spark/ (hybrid sliding-window/full attention, gated residuals, scaled and soft-capped logits), a Bridge plugin, an HF converter and a fixed chat template. scripts/run_muse_spark_64xh200.py (8×8, TP8/PP8/EP8, four 16-GPU engines) and scripts/run_muse_spark_32xb300.py (4×8, TP8/PP4/EP8) replace the radixark/miles:muse-spark-open image; the locked sglang-miles serves the checkpoint and refreshes its derived weight caches after each sync (sgl-project/sglang#49688). Limits: text-only (loading skips the vision and audio towers, the colocated engines serve zeroed towers, and image requests are unsupported), full-parameter only (LoRA is rejected), and --optimizer muon passes --chunked-muon-master-weights, which the locked Megatron-LM does not have (radixark/Megatron-LM#126 is still open), so use the default Adam with NVMe-streamed state. Evidence in this release is the last-4-layer partial on 4×H200 in CI; the full-model numbers on the docs page come from the separate-image era. The docs page still describes the old image and a --vlm option that is not in main (#4759, #4117 @guapisolo).
MiniMax-M3 and M3.1-Flash. A Megatron plugin for the 60-layer MoE with MSA block-sparse attention (miles_plugins/models/minimax_m3/), with a Bridge mapping, HF converter, scripts/run_minimax_m3.py and a BF16 dequantization tool for M3.1-Flash's MXFP8/NVFP4 release. The full MiniMax-M3 ran 10 steps on 8×4 GB300 (TP4/PP4/EP8, colocated, R3): reward 0.52→0.79 with train/rollout log-prob difference 0.0070–0.0096 (PR body). Requires --qkv-format thd, TP ≤ 4, no CP; the indexer stays frozen; the rollout side needs sgl-project/sglang#47612, which is in the locked sglang-miles. M3.1-Flash reuses the plugin on a dequantized copy and is validated on a 4-layer slice only; its full 8×4 GB300 recipe has not been executed (docs, #4581–#4584 @Zhichenzzz).
GLM-5.3-Flash on torchtitan. --train-backend torchtitan --titan-model-name glm5_next trains GLM-5.3-Flash on torchtitan v0.4.0 with FSDP + EP (no TP; PP and CP were exercised on the 4-layer slice only), with FSDP2 CPU offload (--optimizer-cpu-offload) and Adam moments on NVMe (--stream-optimizer-state-to-disk). On 4×8 H200 (FSDP32/EP32, DAPO-Math-17k, R3) AIME 2026 pass@1 went from 28.5% to 57.1% by step 39, where the run was stopped at 41 of 121 steps (PR body, measured on the v0.3.0-era code; the v0.4.0 port was compared against it on the 4-layer slice on 8×H200: log-prob mean absolute difference 6.0e-3, argmax agreement 98.4%). The three 4-layer e2e tests are registered but disabled pending a CI image with GLM-5.3-Flash serving, and the recipe docs still point at radixark/miles:glm53next (docs, #3989, #4539 @Zhichenzzz, #3995, #4439 @HJSang).
DeepSeek-V4 / V4.1. DeepSeek-V4.1-Flash trains LoRA adapters on attention (wq_a, wkv, wq_b, grouped wo_a, wo_b) and the shared expert through native LoRA (scripts/run_deepseek_v4_1.py --train-mode lora; full model colocated on 2×8 H200 with FP8 rollout: log-prob difference 0.02–0.03, raw reward +0.11 from the first to the last 20-step window of a 100-step run, PR body). Chat templates for V4 and V4.1 now render through SGLang's DeepSeekRecipe instead of vendored encodings, which changes rendered prompts in a few cases (see Breaking Changes). The CSA2 candidate selection honors FlashInfer tie policies; defaults are unchanged, and the rollout half of the smaller-index tie contract needs sgl-project/sglang#47960, which is not in the locked sglang-miles (#4205 @yueming-yuan, #4596 @guapisolo, #4740 @zianglih).
Nemotron hybrid models with packed sequences. THD PackedSeqParams now carries total_tokens, so Megatron builds the seq_idx that resets Mamba state at packed-sample boundaries. On Nemotron 3 Super (10 GRPO rollouts) the train/rollout log-prob difference fell from 0.139 to 0.020 (PR body). Applies to every THD model at the cost of one device sync per micro-batch (#4222 @omri-manolo).
Diffusion models (radixark/miles_diffusion). Diffusion RL lives in radixark/miles_diffusion, which has no release tag of its own; this section covers its main from the miles v0.1.0 publish time (2026-08-18) to 841089c5d5 (2026-11-11): 126 PRs from 9 contributors, counted separately from the 400 miles PRs above. Its documentation is ported into this release under docs/diffusion/ as of 2026-10-19 (#4131); pages for the items marked (after the port) are not in this release's docs yet.
| Model | Task | Training | What changed | PRs | Docs |
|---|---|---|---|---|---|
| Krea-2 | T2I | DiffusionNFT, LoRA (2 / 8 GPUs sync, 6 + 2 async); async Flow-GRPO with score centering | Newly supported | #214, #232, #246 @zhihengy, #309, #354 @Rockdu, #298 @HJSang | link |
| Qwen-Image 2.1 | T2I | Flow-GRPO + PickScore, LoRA (1 GPU) | Newly supported (own family qwen_image21) |
#256 @niehen6174, #326 @Rockdu | link |
| MiniMax H3 | T2VA, Ref2VA | LoRA SFT: joint video + audio supervision (8 GPUs, 2-GPU e2e); Ref2VA code and recipe only | New SFT tasks | #311, #312, #303, #295, #296, #292 @Rockdu | guide |
| Stable Diffusion 3.5 | T2I | Flow-GRPO + HPSv2.1; Flow-GRPO + OCR and PickScore mixture; HPSv3 launcher | New recipes (HPSv2.1 and OCR + PickScore verified; the HPSv3 launcher is not) | #237, #241, #239 @Rockdu, #341 @lijiayu0627 | link |
| FLUX 3 Image Klein | T2I | DiffusionNFT + PickScore; Flow-GRPO + HPSv2.1 | Model page only; the family code is not in main |
#317 @Rockdu | link |
Krea-2. Krea-2-Raw trains DiffusionNFT with the OCR reward (LoRA 32/64, synced over CUDA IPC when colocated and over NCCL in the async run; EMA rollout policy; 10-step ODE rollouts, no CFG): the 8-GPU synchronous run moves rollout/reward/raw_mean from 0.28 to 0.94 with eval OCR 0.923, and the asynchronous 6 + 2 run reaches eval OCR 0.875 at about 79 s per rollout (docs page, PR bodies). RMSNorm weights stay fp32 through the all-gather, which separated a flat reward from one matching the reference Flow-GRPO code (#214, #232 @zhihengy, #309, #354 @Rockdu).
Qwen-Image 2.1. A separate qwen_image21 family (Qwen3-VL conditioning, single-stream DiT, RGBA VAE, CFG-distilled) with a rollout patch group that recomputes the text prefix every step, and a 1-GPU PickScore Flow-GRPO recipe (reward about 0.82→0.87 over 200 rollouts on one H200, docs page); needs diffusers 0.41.0 (#256 @niehen6174, #326 @Rockdu).
MiniMax H3 SFT. t2va LoRA SFT encodes the target's soundtrack and feeds both streams to the DiT, --fsdp-supervised-streams chooses which streams add loss, SFT draws one grid index per pair, and a 2-GPU deterministic SFT e2e guards encode, cache and train; the Ref2VA task (ordered image, video and audio references to a video with its soundtrack, trained on transformer_ref) ships as code and an 8-GPU recipe with no curve or e2e yet (#311, #312, #303, #295, #296, #292 @Rockdu).
Features
Multimodal RL in token space. With --multimodal-token-space-processing, Miles ships media file paths instead of pixel tensors, and every trainer rank rebuilds the inputs with SGLang's own token-space processor from the same files and engine config, so training and rollout see the same tokens, mRoPE positions and vision features. Qwen3.5 models train under raw Megatron with a frozen pure-torch vision tower (bitwise with SGLang's on H200), with mRoPE and context parallelism; the trainer reads the engines' on-disk item feature cache instead of re-decoding media (TimeLens Qwen3.5-9B: trainer media wait 8.6→1.0 s per step, PR body), and inputs are rebuilt on a background thread pool as rollout data arrives. Recipes: TimeLens video temporal grounding on Qwen3.5-9B (8×H200, 300 steps, Charades tIoU 0.482→0.573 best) and Qwen3.5-35B-A3B (100 steps, 0.461→0.545), geo3k multi-turn on Qwen3.5-35B-A3B (raw reward 0.67→0.91 over 100 rollouts), geo3k single-turn on Qwen3.8-27B, and Inkling-Small LoRA on geo3k images and AVQA audio (held-out accuracy 55.1→72.8 over 50 rollouts) (docs pages and PR bodies). Engines need --sglang-enable-token-space-processor; the 35B recipes also use --sglang-mm-enable-dp-encoder and the disk preprocess cache, all in the locked sglang-miles. Megatron only; rejects external engines; media passed as bytes rather than paths are written to --shared-media-dir, which must be a shared filesystem on multi-node. The legacy HF-processor path is unchanged and still covered by CI; Inkling multimodal training now requires token space. A new multimodal CI domain runs four e2e files on 4×H200 (TimeLens, AVQA, geo3k, #4240, #4272, #4610, #4633, #4651, #4732, #4742, #4746, #4749 @Rockdu, #4693 @yueming-yuan; SGLang side sgl-project/sglang#46681–#46769, #47903–#48305, #49126).
LoRA adapter checkpoint library. One on-disk adapter format for Bridge, native, torchtitan and Tinker runs: {save}/iter_N/adapter (a symlink to .adapter.<commit_id>/) holds the fp32 masters and Adam moments as a DCP adapter_state/ keyed by HF module, an adapter_parametrization.json, the training state and, when the adapter has a plain-PEFT form, a PEFT export (native canonical GDN adapters have none). A checkpoint saved under one TP/PP/EP/DP layout resumes under another (Qwen3-30B-A3B 2-layer, TP2×EP2 ↔ TP1×EP4, bitwise on B200 per the PR bodies). Sources accepted by --lora-adapter-path: a miles adapter dir, an HF PEFT pair (weights only), Tinker-named exports, legacy per-rank .pt dirs (same layout), Megatron-Bridge PEFT iter_N/ (Bridge mode), and torchtitan DCP step-N/. --lora-adapter-convert-save rewrites any source as an adapter dir. --save on an adapter now needs Adam with fp32 masters, DistributedOptimizer and Gloo, and refuses training fault tolerance and streamed optimizer state (docs, #4608, #4614, #4616–#4620, #4623, #4624 @yueming-yuan).
Multi-LoRA on native adapters, and a wider native LoRA. Native (Megatron raw-mode) LoRA now serves routed adapter slots on dense projections, Qwen3.5/3.6 GDN layers and lm_head, so the Tinker gateway runs with --megatron-to-hf-mode raw (dense models; an 8-GPU multi-tenant e2e covers Qwen3-4B-Instruct). Native single LoRA also gains: Qwen3.5/3.6 GDN adapters (in_proj_q/k/v/z/b/a as independent adapters), an lm_head adapter on every family except Kimi K3, GLM-5.3-Flash and DeepSeek-V4.1 (needs --vocab-size except on Inkling; refused with --save-hf on tied embeddings), DeepSeek-V4.1-Flash attention and shared-expert adapters (scripts/run_deepseek_v4_1.py --train-mode lora, rollout side sgl-project/sglang#45231), and disaggregated engines plus PP>1 over broadcast (validated by hand on Qwen3-0.6B, B200, not in CI). Bridge canonical LoRA splits Qwen3.5/3.6 in_proj into components through the Megatron-Bridge bump, and canonical single LoRA with --experts-shared-outer-loras is accepted. Native multi-LoRA still refuses routed-expert adapters, PP>1 and colocation (docs, #4652, #4662, #4710, #4569, #4578, #4579, #4580, #4458, #4542, #4205 @yueming-yuan).
Tinker gateway on SDK 0.36. The gateway pins tinker==0.36.0 / tinker-cookbook==0.5.9 and rejects other client versions at the first call. New: POST /api/v1/sessions/{id}/finish so ServiceClient.close() frees adapter slots immediately, save_weights_external (official Tinker format, optional age encryption via pyrage), get_external_weights_urls (--tinker-public-url for ingress), ttl_seconds on every save with a sweeper, set_checkpoint_ttl and delete_checkpoint. Canonical MultiLoRA is the gateway default and combines with shared-outer expert LoRA; per-request gradient accumulation fixes a 25% expert-gradient error at EP2. MiMo-V2.6 tenants run through scripts/run_mimo_v2_6_flash_tinker.py with --rollout-precision bf16|fp8|fp8-mxfp4 (rollout side sgl-project/sglang#45939); in the launcher check with fp8-mxfp4 engines, four DAPO-Math tenants' accuracy rose from 0.33 to 0.69 over 10 steps at 184 s/step, and on the full model on 16×H200 GSM8K accuracy went from 0.60–0.64 to 0.92–0.96 (PR body). Agentic Tinker training ships as a standalone tinker_agent capture-proxy client and a Harbor Terminal-Bench 2 recipe on Qwen3.6-35B-A3B (8×H200, 4 train + 4 sample) (docs, #4611, #4565, #4625, #3905, #3899, #4480, #4507 @yueming-yuan, #4326 @guapisolo, #3682, #3684 @yushengsu-thu).
torchtitan LoRA and HTTP adapter transport. --train-backend torchtitan trains one LoRA on Qwen3's linear projections (--lora-type lora only), syncs the adapter to SGLang as a PEFT adapter, and checkpoints through the adapter library (Qwen3-4B, 4×H200, TP2: reward 0.22→0.41 over 4 rollouts, ~3.0 s per sync, PR body). --update-weight-transfer-mode http-lora hands a single adapter to external engines over HTTP, either streamed through the engine weight-update session (--http-lora-ship tensors, needs sgl-project/sglang#48241) or as a versioned PEFT dir loaded with upsert (path, needs sgl-project/sglang#47135); a B300 trainer fed 8×MI355X engines a 4.1 GB GLM-5.3 adapter in 84 s per sync (PR body). Refused: colocate, multi-LoRA, PD disaggregation (#4092 @HJSang, #4620, #4624 @yueming-yuan, #3690 @Weyaxi).
TPU. The torchtitan trainer runs on TPU v7x through torch_tpu, next to an external sglang-torchtpu engine on the same host (4 trainer devices + 4 engine devices on one 2x2x1 host), with weights moved over ICI by the new --update-weight-transfer-mode ici. Launchers under scripts/tpu/ cover Qwen3-4B, Qwen3-30B-A3B (CPU-offloaded Adam) and a 4-layer GLM-4.7-Flash; --compile-warmup compiles every pass on synthetic batches while the engine starts (step-0 log-prob + train 1.8 s + 14 s instead of 1265 s + 592 s on Qwen3-4B, PR body); --titan-fsdp-gather-once and --titan-fsdp-reduce-dtype are platform-neutral opt-ins; batched ragged paged attention raised rollout throughput about 1.38× on the Qwen3-4B bench recipe (PR body). Evidence is smoke (3 rollouts) and bench (4 rollouts) runs on one host; the only long runs are one 543-step epoch with an fp8 KV cache, which collapsed, and a bf16-KV epoch still running at merge, which is why TPU engines default to a bf16 KV cache. The radixark/miles:*-tpu image (docker/build.py --variant tpu) is built by manual dispatch only; CI runs a CPU contract test against the pinned sglang-torchtpu (README, #4213, #4214, #4215, #4218, #4226, #4265, #4268, #4597, #4598, #4656, #4720, #4724 @HJSang, #4589, #4713, #4730, #4758 @Zhichenzzz).
Dataset mixing. --rollout-data-yaml-config lists several training datasets; every rollout step draws from them in a fixed samples_per_step ratio (smooth weighted round-robin, exact even when fully-async asks one prompt at a time). Each dataset carries its own position, epoch and seeded shuffle, an optional sha256 pin, prompt/response/sequence caps, generate or agent function, free-form agent block and reward spec (custom_rm_path / rm_type); metrics are logged per dataset under rollout/dataset/<name>/. Positions are checkpointed and a mismatching resume fails unless --rollout-data-reset-on-load restarts the datasets while keeping model, optimizer and step counters. Requires --num-rollout; refuses --partial-rollout, --num-epoch and a custom --data-source-path (docs, #4492, #4553–#4561, #4613, #4669, #4491, #4490 @Shi-Dong).
Agentic recipes and sandboxes. Novita Agent Sandbox joins E2B, Daytona and Modal as a sandbox provider for the Harbor and OpenEnv recipes (pip install -e '.[novita]', NOVITA_API_KEY). A text-to-SQL multi-turn recipe trains Qwen3.6-35B-A3B on one 8×H200 node with read-only execute_sql_query tool calls: BIRD-Platinum validation accuracy 61.3→90.2 and Arcwise 52.6→90.0 over 160 steps (docs page). A tau2-bench recipe runs unmodified tau2 simulations through the session server on Qwen3.6-35B-A3B; its README states it does not yet improve pass rate (docs, text-to-SQL, #4435 @CeerDecy, #4674, #4675 @Zhichenzzz, #4711 @guapisolo).
GRPO length penalties and bounded staleness. --length-penalty-coeff (default 0, off) shapes correct trajectories' rewards by their length relative to the shortest correct sibling in the group; the raw reward is still logged. --async-submit-within-staleness applies --max-weight-staleness when the fully-async producer submits a group, so no group is generated only to be dropped: on Qwen3-4B GRPO GSM8K with 4×MI350X, stale groups fell from 39–41% to 0 and step time from 11.7 to 9.8–10.0 s (PR body). Requires --fully-async, --update-weights-interval 1 and one optimizer step per rollout (docs, #4430–#4432, #4441 @Shi-Dong, #4270 @Arist12, #4733 @HJSang).
NIXL weight transport and router backends. --p2p-transport nixl (default mooncake) writes P2P weight updates over NIXL; the engines start with --sglang-remote-instance-weight-loader-start-seed-via-nixl and need sgl-project/sglang#35712 (in the locked sglang-miles). On 4×H200 (Qwen3-30B-A3B, one TP2 engine) steady-state updates take 3.0–3.1 s over Mooncake and 3.1–3.5 s over NIXL (PR body); the nixl package comes from the SGLang base image. --router {sglang,miles} (default sglang) replaces the use_miles_router switch behind a RouterBackend that owns launch arguments, worker environment, membership and routing-key headers; the Miles router now serves sgl-router's /workers API (docs, #2721 @amitw-nv, #4039 @yueming-yuan, #4317, #4320–#4322 @HJSang, #4527 @AndyDai-nv).
FSDP on FlashAttention-4. --attn-implementation flash_attention_4 runs the FSDP backend on the FlashAttention that SGLang already installs, so an FSDP pip install trains without a flash-attn build; adapters for Qwen3, Qwen3-MoE, Qwen3-VL and GLM-4.7-Flash pass packed-document boundaries to HF attention. Default stays flash_attention_2 (docs, #4316 @HJSang).
Diffusion: DiffusionNFT aligned with the reference. The old policy is the EMA (--loss-type nft requires --use-ema --rollout-weights ema), train timesteps come from a pluggable strategy (default drop_random_steps with 0 dropped steps; the recipes drop one), each pair reads its own sample's sigmas, the KL term is logged as kl_loss, and the engine returns only the final x0 for NFT requests (#222, #266, #272, #280, #281 @Rockdu).
Diffusion: Flow-GRPO on recorded sigmas, with score centering. SDE steps are scored with the sigma, next sigma and sigma_max the rollout recorded, through a FlowSdeStepBackend that replaces the train-side diffusers scheduler. --diffusion-score-centering (experimental, after the port) subtracts a closed-form clipped-ratio center so the policy-gradient step stays unbiased under train/rollout mismatch: on Krea-2 async Flow-GRPO with an fp8 rollout (4 + 4 GPUs, 300 rollouts, two seeds) the uncorrected runs collapsed to zero reward near rollouts 200–270 while the corrected ones kept the clip fraction at or below 1.7% and ended at mean rewards 0.902 and 0.842 (PR body). It requires --diffusion-debug-mode and an SDE rollout and has CPU tests but no e2e standard (#278, #279, #320 @Rockdu, #298 @HJSang, #380 @Rockdu).
Diffusion: DMD2 distillation (after the port). --loss-type dmd2 adds a critic trainer group before the actor: the rollout stops at a random step, the critic (a fake-score model initialized from the teacher) trains on the student's samples, scores them against the frozen teacher's forward, and hands the distribution-matching gradient to the student; a MiniMax H3 4-step recipe on 2×8 GPUs ships with eval media and DMD2 metrics. The engine side (stop at a requested step, skip the VAE decode) is on sglang-miles since 2026-11-11 (sgl-project/sglang#49440, #49143); the recipe has not run at its 16-GPU scale and has no e2e yet (#331, #333, #334, #343, #372 @zhihengy).
Diffusion: SFT data path. train_sft.py prefetches the next encode while the current step trains; SFT media decodes through ffmpeg into raw uint8 clips with a lazy MediaSource; rows carry ordered conditions and a target media dict; encoders receive the whole sample through encode_sample; the noise grid is built train-side from a per-stream --fsdp-flow-shift; LoRA checkpoints export to loader-ready safetensors (#248, #200–#204, #206, #287–#290, #294, #276, #277, #286 @Rockdu).
Diffusion: engine from sglang-miles. The rollout engine builds from SGLang's sglang-miles branch (adapted to v0.5.22 and v0.5.23), rollout patches are applied through SGLang's scheduler bootstrap again, only the SDE window's latents are returned by default, the router streams responses (RSS 34 GiB→144 MiB at 16-way concurrency), and --diffusion-quality (default exact) pins the engine's numerics (#257, #315, #350, #284, #216, #199, #198 @Rockdu).
Diffusion: async training. train_diffusion_async.py submits the next rollout before training the current one and updates weights with no generation in flight; adapters or full weights move over NCCL to separate rollout GPUs; --async-exact-resume restores the exact EMA on resume. Validated on Krea-2 DiffusionNFT with 6 + 2 GPUs and a 4-GPU proxy e2e (#232 @zhihengy, #298 @HJSang).
Diffusion: rewards. --rm-type hpsv3, openai_api / custom_api rewards configured from a file, a shipped weighted mixture reward with per-reward scales and monitors, and per-reward colocation flags (#341 @lijiayu0627, #244 @JingwenGu0829, #237 @Rockdu).
Optimizations
Replay outputs through the output store. With --use-output-store, SGLang writes each response's routed experts, indexer top-k and sampling mask into the run's Mooncake store and returns a reference; the trainer reads the objects back in one call per shard and rebuilds the same training columns (6.3 MB per 2048-prompt/1024-output response no longer travels inline, PR body). Objects are deleted after the rollout trains and leftovers are swept at run end. Off by default; needs --object-store-backend mooncake, the radix_transfer_plugins package the images now install, and the Miles router or a forwarding sgl-router; rejected with Kubernetes, PD disaggregation, external engines, multi-policy, --multimodal-keys, --dump-details and a custom fully-async data buffer. Twelve R3 e2e tests were switched to the store (eleven remain registered in the release tree) with unchanged steady-state step times (docs, #4236, #4242–#4244, #4264, #4267, #4478, #4509 @guapisolo; SGLang side sgl-project/sglang#42647, sgl-project/sglang#46288).
DSA, indexer and GDN kernels. The DSA sparse-attention backward runs on cuDNN in the tuned rows for sm90/sm100 (GLM-5 1.05–4.8× on H200 and 1.9–7.4× on GB300 over the TileLang backward; DeepSeek-V4 up to 9.6×, PR body), the indexer top-k defaults to an exact canonical selection with Triton scoring (--miles-dsa-topk-backend canonical; torch restores v0.1.3), and Qwen3.5 / Qwen3-Next GDN layers use TE's fused input norm and a single in_proj GEMM (--linear-attention-input-norm te; hf restores the eager norm). End to end on 4×GB300 (15 rollouts) the GLM-5 4-layer actor train step fell from 16.3 to 12.1 s; the new defaults differ from the old paths at bf16 level (log-prob relative difference 3e-7 to 9e-6). On MI350X, gfx950 gets its own DSA tuning rows, fixing a dKV relative error of 0.40 (#4001 @Zhichenzzz, #4426 @XinyuJiangCMU).
Smaller things that add up. The Megatron argument parser is built once per command line (multi-policy e2e 18,990 s → 3,134 s on 4×H200, PR body) (#4328 @HJSang). The fp8 quantizer decides each prefix's skip once per config (26.3 → 23.2 s per sync on GLM-5.2 fp8, 2×8 H200) (#4447 @yueming-yuan). Multi-LoRA Tinker trainers split packed response logits once (≈83 → 73.5 GiB peak on Qwen3.6-35B-A3B 65K replay) (#4343 @yueming-yuan). A colocated LoRA run with --offload-train-target disk backs up only the trainable parameters when the engines keep the base (#4291 @yueming-yuan). --stream-optimizer-state-page-cache-mb caps the host page cache of NVMe-streamed Adam state (#4143 @guapisolo). Megatron-Bridge plugins load lazily instead of on megatron_utils import (#4562 @yueming-yuan).
Diffusion: rollout data plane (experimental, after the port). With --use-output-store, the engine writes each step's latents, debug and condition tensors and the decoded output into the run's Mooncake store and returns references; the trainer reads its shard back before training and objects are released after the rollout trains (the decoded output alone was 79% and 93% of the response for SD3.5 NFT and LTX-2.3, PR body). Off by default; needs --output-store-mooncake-master and the radix_transfer_plugins package the image installs; six RL e2e tests run through it bitwise (#359, #360, #361, #363 @zhihengy, #370, #373, #376 @Rockdu; SGLang side sgl-project/sglang#42647, #48959, #49693, #49877).
Diffusion: weight sync through miles' HF weight iterator. All five sync modes (full or LoRA, colocated IPC or NCCL) stream through one iterator with per-parameter transforms and atomic groups (#304 @Rockdu).
Refactors
Typed, immutable configuration. Every Miles-owned CLI argument is declared as an Annotated[..., Arg(...)] field on a pydantic config under miles/utils/args/configs/; backend-owned arguments (Megatron's parser, the FSDP/torchtitan dataclasses) sit in their own namespaces under it. parse_args() returns an AllConfig whose fields are immutable after parsing, apart from declared mutable scopes, and which rejects parsed fields without an owner. Trainer, rollout and orchestrator each receive their own typed slice (TrainerConfig with a backend-owned namespace, OrchestratorConfig), SGLang and router settings are resolved once at the parse boundary, and checkpoint-load inputs travel as call-scoped runtime inputs. Custom functions declare their options with fn.config_class = MyConfig(BaseConfig) and receive them through an immutable flat view; --custom-config-path still works with a deprecation warning. A snapshot CI compared pre- and post-refactor configurations across the whole test matrix before being switched off. The series is about 160 PRs, all by @fzyzcjy, with the remaining PRs in these ranges listed under their own headings (the ranges span only the PRs merged in this release) (#3469–#3576, #4344–#4425; user-visible consequences under Breaking Changes).
Pure wheel, torchtitan v0.4.0. miles builds as one py3-none-any wheel that imports the FSDP and torchtitan backends without Megatron (a wheel-smoke CI hides Megatron and imports both); log probabilities use one plain-torch vocab-parallel cross entropy on every backend (bit-identical to Megatron's unfused kernel, last-ULP differences from the fused one, ~9% slower at TP=1 on H200). The torchtitan backend moves to v0.4.0 on torch 2.14 with backports in compat.py; routing replay on static-shape platforms stages ids on the host, which also removes one device sync per MoE layer on GPU; losses are written once against a response layout (#4308, #4315, #4214, #4218 @HJSang, #4537 @Zhichenzzz).
Service ports outside the ephemeral range. Each node derives its service-port range from the kernel's ip_local_port_range (20000–32767 on a stock kernel instead of 20000–65535) and Ray workers get --min-worker-port 0 --max-worker-port 0 when Ray's default 10002–19999 overlaps the ephemeral range, ending the bind races seen on long runs (docs, #4199 @bobboli).
Diffusion: trainer. Frozen reference and teacher models through one loader, EMA as an optimizer checkpointed under ema/, pinned-memory sleep/wake (sleep 2.77→1.55 s, wake 0.94→0.22 s on SD3.5 OCR GRPO, 2×H200), loss hooks selected by (loss type, role) with a frozen context and the teacher forward in the actor, several colocated trainer groups in the sync driver (#263, #267, #268, #269, #273, #265, #367, #368, #372 @Rockdu, #338, #329, #330, #335, #342 @zhihengy).
Diffusion: release infrastructure (after the port). Versioned release workflows, a dependency lockfile and a release runbook ported from miles; nothing has been cut yet (#369 @Rockdu).
Full release notes by category below; breaking changes are near the end.
Model Support & Fixes
- Muse Spark Open model code, Bridge plugin, HF converter, fixed chat template, 64×H200 / 32×B300 launchers and a last-4-layer partial CI; MFU sizing for its output gate and latent experts: #4759, #4117 @guapisolo
- MiniMax-M3 plugin (MSA block-sparse attention, QSA tile map and Triton block-max indexer scores), Bridge mapping, converter, launcher and docs; MiniMax-M3.1-Flash recipes on a BF16 dequantized copy: #4581, #4582, #4583, #4584 @Zhichenzzz
- GLM-5.3-Flash (
glm5_next) on torchtitan: 4-layer FSDP+EP+PP+CP, full model with CPU offload and NVMe Adam, port to torchtitan v0.4.0, transformers 5.19indexed_attentionlayer type: #3989, #4539 @Zhichenzzz, #3995, #4439 @HJSang - DeepSeek-V4.1-Flash native LoRA on attention and the shared expert, with two
--load-from-hffixes: #4205 @yueming-yuan; CPU test fix after the target-resolution change: #4482 @HJSang - Kimi-K2.5 native LoRA no longer crashes the SGLang engines after the first weight update and LoRA load: the DeepSeek/Kimi MHA cached-prefix paths applied the LoRA-wrapped
kv_b_projto prefix rows with the batch's extend-token segments (an out-of-bounds write surfacing asCUDA_ERROR_ILLEGAL_ADDRESSin flashinfer's ragged attention); the lockedsglang-milesdisables the chunked prefix cache when LoRA targetskv_b_projand asserts dense LoRA input rows. Verified by a rerun oftests/e2e/megatron/model_scripts/test_kimi_k25_2layer_lora_native_ci.pywith no engine crash (sgl-project/sglang#49848 @Rockdu). The same rule rejectskv_b_projadapters on DSA engines, so the GLM-5 / 5.1 / 5.2 default LoRA targets and the GLM-5.2 native and Daytona recipes now leavekv_b_projfrozen (#4786 @HJSang); see Breaking Changes - DeepSeek-V4 / V4.1 chat templates rendered through SGLang's
DeepSeekRecipe(vendored encodings removed): #4596 @guapisolo - DeepSeek-V4.1 candidate selection honors FlashInfer tie policies; replay once per forward; score-tensor chunking bounded to 0.5 GiB: #4740 @zianglih
- Qwen3-VL MoE under Bridge mode trains the language model and ViT again instead of only the vision mergers: #4331 @Rockdu
- Mamba state reset at packed-sample boundaries in THD (Nemotron-H / Nemotron 3): #4222 @omri-manolo
- Inkling:
--ffn-hidden-sizeis the dense MLP width; variable-sequence fallback restored; docs point atradixark/miles:latest: #4605, #4687 @yueming-yuan, #4417 @fzyzcjy - Kimi-K3.1 W4A16 LoRA recipe page (separate image): #4665 @guapisolo
Multimodal (Token Space)
- Media references for token space; training inputs built in token space; Qwen3.5 multimodal under raw Megatron (mRoPE, CP, frozen pure-torch vision tower); TimeLens video RL with background input rebuild: #4240, #4272, #4610, #4633 @Rockdu
- Inkling image and audio RL on the token-space processor; AVQA audio example and Inkling-Small LoRA profile: #4732 @Rockdu
- Trainer reuses the rollout engines' on-disk item feature cache: #4651 @Rockdu
- Routed-experts replay sized from the engine's prompt and completion token counts on expanded multimodal prompts: #4693 @yueming-yuan
- geo3k token-space recipes for Qwen3.5-35B-A3B, Qwen3.8-27B and Inkling; Qwen3.5-35B-A3B TimeLens profile;
multimodalCI domain with four e2e files: #4742, #4746, #4749 @Rockdu - Snapshots re-recorded for the engines' token-space and cache server args: #4535, #4721 @Rockdu
LoRA & Multi-LoRA
- Backend-agnostic adapter checkpoint library; Bridge adapter slots and PEFT
--save-hf; Bridge and native single LoRA save/resume through adapter dirs (rank.ptwriter dropped); Megatron-Bridge PEFTiter_N/import;--lora-adapter-convert-save;tools/convert_lora_adapter_to_peft.py: #4608, #4614, #4618, #4619, #4623 @yueming-yuan - Tinker sampler exports and
save_state/load_stateon the adapter library, resolved by commit; external weights download, checkpoint TTLs and deletes; session finish endpoint; Tinker SDK 0.36 wire schema: #4616, #4617, #4625, #4565, #4611 @yueming-yuan - Multi-LoRA on native adapters:
AdapterSlotscompute members, routed native slots, Tinker gateway in raw launcher mode with padded-vocab HF load: #4652, #4662, #4710 @yueming-yuan - Canonical MultiLoRA (Tinker default), shared-outer expert LoRA in Tinker, canonical single LoRA with shared-outer experts, adapter-layout CI matrix, per-request gradient accumulation fix: #3905, #3899, #4542, #4507, #4480 @yueming-yuan
- Native LoRA: Qwen3.5/3.6 GDN adapters,
lm_headadapter on every family except Kimi K3, GLM-5.3-Flash and DeepSeek-V4.1, per-backend--lora-typedefault withcanonical_lorarequired on native, explicit adapter set for the exporter,in_proj_b/adropped from theattngroup with split-fused-module refusal, Megatron-Bridge bump for canonical GDNin_projand gated-q adapters: #4569, #4578, #4574, #4566, #4580, #4579 @yueming-yuan - Native LoRA in disaggregated mode and PP>1 for native and Bridge single LoRA; multi-LoRA no longer silently overrides
--megatron-to-hf-mode;--save,--save-intervaland--save-hfrejected in Tinker mode: #4458, #4563, #4564 @yueming-yuan - torchtitan LoRA with adapter weight sync; adapter checkpoint members and
probe_torchtitan_row_mapping; save and resume through adapter dirs; LoRA on torchtitan v0.4.0's fused projections: #4092 @HJSang, #4620, #4624 @yueming-yuan, #4538 @Zhichenzzz - MiMo-V2.6 multi-LoRA through the Tinker gateway with
--rollout-precision: #4326 @guapisolo - Pinned rollout adapter on SGLang attention DP (multi-LoRA refused there): #3747 @yushengsu-thu
- Colocated LoRA host backup limited to trainable params when the sync skips the base; packed response logits split once with profiler work units; deterministic HF target ordering; Bridge plugins loaded with
AutoBridge: #4291, #4343, #4562 @yueming-yuan, #4395 @fzyzcjy - GLM-5 / 5.1 / 5.2 default LoRA targets and the GLM-5.2 native and Daytona recipes leave
kv_b_projfrozen: #4786 @HJSang - Multi-LoRA training service overview page: #4476 @yueming-yuan
RL Algorithms & Training
- Group-relative correct-answer length penalties (
--length-penalty-coeff,--length-penalty-min-tokens), shared prompt/trajectory grouping for reward processing, per-session logging of counted tokens and applied penalties: #4430, #4431, #4432, #4441 @Shi-Dong - Megatron log probs computed in train mode unless the model has dropout: #4107 @yueming-yuan (with radixark/Megatron-LM#136, #137)
- LR scheduler stepped on resume only when the checkpoint did not restore it; geometric TIS weights normalized over sequences; sampling masks fetched for
custom_lossunder support replay: #4530, #4528, #4529 @clumsy - Reward hub: gpqa rewards only a single committed option;
boxed_prefix fixed for math, deepscaler, gemma_math and dapo; sympy power guard covers**;reward_spec.custom_rm_pathfollows the--custom-rm-pathhook and batched-group rules: #4442, #4443, #4454, #4490 @Shi-Dong distill_qwenmask keeps the full prompt context and trains EOS; SFTloss_maskaligned withresponse_lengthfor fully masked samples;mask_truncated_zero_rewardhandles nested session-server groups: #4444, #4456, #4455 @Shi-Dong- Rollout top-k candidates checked against the trainer in true-on-policy CI: #4109 @HJSang
- Plain-torch vocab-parallel cross entropy for log probs on every backend: #4315 @HJSang
Rollout & Session
- Dataset mixing:
--rollout-data-yaml-config(validated list, weighted mixing source, checkpointed positions, sha256 pins, per-dataset metrics, generate and agent functions, agent block, sequence limits,max_response_len),--rollout-data-reset-on-load: #4492, #4553, #4554, #4555, #4556, #4557, #4558, #4559, #4560, #4561, #4613, #4669 @Shi-Dong Sample.max_response_lenlets a sample lower its own response budget: #4491 @Shi-Dong- Output store: merge plan (
Seg,Segments,StorePlan), plan carried on samples, session reply and train dict, plans built from rollout responses, replay columns materialized on the trainer, objects deleted once unread,--use-output-store, most R3 e2e tests through the store with a leftover sweep, custom fully-asyncDataBufferrejected with the store: #4236, #4242, #4243, #4244, #4264, #4267, #4478, #4509 @guapisolo --max-seq-lenreplaces--rollout-max-context-len; warning that it does not stop an agent under the session server: #4699, #4690 @guapisolo--rollout-shuffleon by default: #4463 @Shi-Dong- Bounded staleness when the fully-async producer submits a group (
--async-submit-within-staleness); orchestrator-side check dropped to fix multi-policy startup: #4270 @Arist12, #4733 @HJSang - GLM stop-token replacement gated behind
--allow-stop-token-replace: #4653 @guapisolo - Last kept turn marked
TRUNCATEDwhen later turns are dropped: #4457 @Shi-Dong - Router: launch command, membership API and routing-key headers behind a
RouterBackend;--routerselects the router; worker environment delegated to the backend;router_dp_awarefor attention DP in either spelling; serving registration behind a registry: #4317, #4320, #4321, #4322 @HJSang, #4527, #4147 @AndyDai-nv, #4313 @yueming-yuan - Dynamo router foundation (parts 1–4 of 13: typed
--dynamo-config, frontend and sidecar argument rendering, sidecar listeners); not launchable yet: #3890, #4700, #4701, #4702 @AndyDai-nv - SGLang server args whose
SUPPRESSdefault leaves the dest unset are skipped: #4691 @yueming-yuan
Agentic & Environments
- Novita Agent Sandbox as a sandbox provider for Harbor and OpenEnv: #4435 @CeerDecy
- Text-to-SQL multi-turn recipe on Qwen3.6-35B-A3B; tau2-bench agentic RL example with cut-off and failed-episode classification: #4674, #4675 @Zhichenzzz, #4711 @guapisolo
- Agentic training through the Tinker gateway: capture-backed proxy client and a Harbor Terminal-Bench 2 recipe: #3682, #3684 @yushengsu-thu
- Examples migrated from legacy rollout functions to the class-based
GenerateFnInputinterface; immutable multi-agent configuration separated from runtime state; metrics aggregated from v2 session samples: #4415, #4398, #4420 @fzyzcjy
Weight Update
- NIXL as a second P2P transport (
--p2p-transport nixl); sharedP2PTransportwith native Mooncake calls: #2721 @amitw-nv, #4039 @yueming-yuan - HTTP-LoRA transport for external rollout engines: #3690 @Weyaxi
- Every rank of a multi-node engine gets node 0's info-bootstrap port; fp8 quantizer skip decided once per config; replicated Mooncake object-store transfers retried: #4448, #4447 @yueming-yuan, #4402 @fzyzcjy
--saveevent directory no longer turns on engine weight checksums; checksums computed only when checked: #4319 @yueming-yuan, #4551 @clumsy- TPU weight sync over ICI for a trainer and engine sharing one slice: #4589 @Zhichenzzz
TPU
- torchtitan trainer on TPU (
torch_tpu): platform capabilities, length buckets, per-block compile, Pallas splash attention; losses written once against a response layout with padded rows on TPU; routing replay without dynamic shapes: #4213, #4214, #4218 @HJSang docker/Dockerfile.tpuon sglang-torchtpu's release image; torch_tpu with job-local worlds; sglang-torchtpu pinned past #1825 and its patches dropped: #4268, #4598, #4656, #4720 @HJSang--compile-warmup,--titan-fsdp-reduce-dtype,--titan-fsdp-gather-once; bf16 KV cache by default on TPU; sglang-torchtpu API contract test on a CPU runner: #4724, #4215, #4265, #4597, #4226 @HJSang- Launchers for Qwen3-30B-A3B and GLM-4.7-Flash (4-layer); batched ragged paged attention for rollouts; MoE launchers' weight sync fixed: #4713, #4730, #4758 @Zhichenzzz
Low Precision & Kernels
- cuDNN DSA backward, exact canonical indexer top-k with Triton scoring, fused GDN input norm and projections: #4001 @Zhichenzzz
- gfx950 DSA tuning rows: #4426 @XinyuJiangCMU
miles-ext-kernelssubmodule scaffold behind--use-miles-ext-kernels(off; not installed in the image): #4274 @yueming-yuan- NVFP4: negative zero cleared in exported E2M1 codes; W4A16 rollout and QDQ design referenced from the docs: #4766 @nanjiangwill, #4340 @zianglih
Memory & Offload
- Host page cache of NVMe-streamed optimizer state capped with
--stream-optimizer-state-page-cache-mb: #4143 @guapisolo - torchtitan:
--optimizer-cpu-offload(renamed from--fsdp-cpu-offload) and Adam moments on NVMe (--stream-optimizer-state-to-disk) for full-parameter GLM-5.3-Flash: #3995 @HJSang - Disk offload target dropped automatically under
--debug-rollout-only: #4123 @guapisolo
Launchers & Configuration
- Service and Ray worker ports kept out of the host's ephemeral port range: #4199 @bobboli
- Megatron argument parser built once per command line; same-second run ids no longer collide: #4328, #4472 @HJSang
- Torch profiler honors
--profile-ranksand the collect flags: #4128 @guapisolo - FSDP on FlashAttention-4 through packed-document boundaries (
--attn-implementation flash_attention_4): #4316 @HJSang - Pure wheel importing FSDP and torchtitan without Megatron; Python ≥ 3.11; wheel-smoke CI: #4308 @HJSang
- torchtitan bumped to v0.4.0 on torch 2.14: #4537 @Zhichenzzz
- Configuration refactor consequences for launch commands:
--miles-router-max-connectionsunset means no cap;--no-load-optim/--no-load-rngon every backend; external engines need not sum to--rollout-num-gpus: #4383, #4394, #3560 @fzyzcjy
Metrics
rollout/raw_rewardreported before dynamic sampling filters (rollout/raw_reward_filteredfor the retained batch): #4486 @guapisolo- Rollout step logged with train metrics: #4552 @clumsy
- MFU sized per layer from the HF config (compressed and sparse-indexed attention included), tracked in the CI metric history;
dsa_indexer_loss_coeffread off the Megatron backend: #4096, #4097, #4098 @Zhichenzzz, #4474 @HJSang
AMD / ROCm
- ROCm images on the SGLang v0.5.23 ROCm bases with expanded MI350 nightly coverage; ROCm CI on the ROCm 10 MI35X image; ten more MI350 nightly tests and three fast-GPU kernel tests; torch_memory_saver 118b09b: #4599, #4510, #4325, #4318 @XinyuJiangCMU, #4514 @michaelzhang-ai
- ROCm PR CI on Spur MI355X runners;
tml-renderersandopenai==2.8.0in the ROCm image; GLM-5.2 launcher keeps the gfx950 fused DSA indexer off so weights update; Verifiers venv fix: #4760, #4670, #4626, #4672, #4660 @michaelzhang-ai - gfx950 DSA tuning rows (see Kernels): #4426 @XinyuJiangCMU
Release CI does not run the ROCm suite; AMD evidence is local MI355X / MI350X runs and the external nightly.
Refactors
- Configuration refactor (about 160 PRs @fzyzcjy): typed
Annotatedargument declarations andAllConfig(#3469–#3478, #3490–#3494, #4361, #4367–#4369); trainer backend namespaces andTrainerConfig(#3487–#3489, #3500–#3518, #3537–#3539, #4363, #4393, #4394); SGLang and router configuration resolved at the parse boundary (#3481–#3486, #4362, #4366); typed custom-function configuration and flat hook views (#3519–#3532, #4364, #4413, #4414); explicit dynamic attribute access (#3495–#3497, #4401); worker specifications, static and dynamic Kubernetes pools, pod metadata and payload transport (#3534–#3556, #3561, #4371, #4381, #4382, #4418, #4422); scaling configuration and observed inference topology (#3558–#3575, #4365, #4376, #4399, #4421);OrchestratorConfig-driven training and trainer controllers built from runtime inputs (#4372–#4380); W&B run id reused across launcher relaunches (#4370); call-scoped checkpoint load inputs (#4405–#4410, #4419); HF→torch_dist converter on typed configuration withCONVERSION_VERSION6 (#4396, #4397); configuration-snapshot CI built, used to validate the refactor, then disabled (#4344–#4360, #4384, #4423–#4425); W&B credential digests normalized when loaded from the environment (#4416) - Orchestration and fault-tolerance internals, and their Kubernetes test harness: #4403, #4404, #4411, #4412, #4323, #3576, #4385–#4392, #4400 @fzyzcjy
- Multi-hardware rule for device-agnostic code: #4464 @HJSang
CI & Testing
- CUDA image installs SGLang's
radix_transfer_plugins; reinstalled after selecting SGLang in CI: #4329, #4658 @guapisolo - Loss snapshots read from
radixark/miles-wheelsat a pinned commit; transient lookups retried in build-wheels; hourlysglang-mileswheel builds: #4511, #4609 @HJSang (#4311 landed through #4316) - Eight-GPU tests rebalanced between H200 and H100;
test_dumperand GLM-4.7-Flash native LoRA moved to H200; GPU labels forbidden on CPU runners; label-command access for michaelzhang-ai and bingxche: #4332, #4339, #4761 @guapisolo, #4734 @yueming-yuan, #4673 @michaelzhang-ai - Snapshots record Megatron's
moe_permute_accumulation_in_fp32andmoe_expert_fake_quant; fixture and registration fixes; flaky-test fixes (GLM FP8 fully-async reward tolerance, deterministic sample filter test, prompt-keyed agentic expectations, raw reward tracked not gated for Kimi K3 4-layer, session identity projected out of parity comparisons): #4485, #4460, #4595 @guapisolo, #4744, #4440, #4531, #4494 @HJSang, #4655, #4591 @Zhichenzzz, #4576 @Rockdu - Version bump to 0.2.0: #4774 @Rockdu
Docs Updates
- New pages: MiniMax-M3, MiniMax-M3.1-Flash, Kimi-K3.1, TimeLens video, AVQA audio, Text-to-SQL, Mixing datasets, Multi-LoRA training service, Harbor + Tinker, Tinker agent, Length penalties.
- Updated pages: LoRA, geo3k (token space), Inkling (multimodal RL,
radixark/miles:latest), DeepSeek-V4.1 Flash (LoRA, tie policies), MiMo-V2.6-Flash (multi-LoRA), training backends (torchtitan GLM-5.3-Flash, TPU flags, FlashAttention-4), fully async (bounded staleness), launch script (ports), monitoring (raw_reward, router/workers, profiler), sandbox providers (Novita), p2p weight transfer (NIXL), Mooncake rollout transfer (output store), disk offload (page cache), CLI reference, customization, train-infer mismatch helper (explicit flags instead of a YAML), contributor guide (multi-hardware rule).
Diffusion (radixark/miles_diffusion)
Diffusion RL lives in radixark/miles_diffusion, which has no release tag of its own; this section covers its main from the miles v0.1.0 publish time (2026-08-18) to 841089c5d5 (2026-11-11): 126 PRs from 9 contributors, counted separately from the 400 miles PRs above. Its documentation is ported into this release under docs/diffusion/ as of 2026-10-19 (#4131); pages for the items marked (after the port) are not in this release's docs yet.
- Models, SFT tasks and recipes: see the Diffusion rows and paragraphs under Highlights › Model Support.
- Algorithms, rollout engine, data plane, async training, rewards and trainer: see the Diffusion paragraphs under Highlights › Features, Optimizations and Refactors.
- Dependencies and image: The rollout engine builds from
sgl-project/sglangbranchsglang-milesat image build time (wasmain;SGLANG_DIFFUSION_COMMITpins a commit);radix_transfer_pluginsandmooncake-transfer-engine==0.3.13.post1for the data plane; diffusers0.41.0(was a git pin) and transformers5.12.1(was 5.5.4);open_clip_torchand anhpsv2git pin for the HPS alignment tests;imageio-ffmpegfor eval audio (#257, #363, #373, #326, #194, #237, #343). No versioned image exists;docker-build.ymlpublishesradixark/miles_diffusion:dev-cu129-sglang-main-<date>andlatest(one variant, CUDA 12.9, x86_64). - Contributors (miles_diffusion): @charliechenye @GuanqiHe @HJSang @JingwenGu0829 @lijiayu0627 @niehen6174 @OnePunchMonk @Rockdu @zhihengy.
Dependencies
- SGLang:
sglang-mileslocked atfd017154ad(v0.1.3:3a938213c1). The Miles patch stack is rebased onto SGLang v0.5.23 (#4548 @Rockdu): 96 commits on top of the release, up from 51 on v0.5.22. New in the stack since the v0.1.3 lock:- Output store: radix transfer plugins, direct reads into caller memory, byte-range reads (sgl-project/sglang#42647, #46288, #47320); for diffusion RL, a tensor stash in
radix_transfer_pluginswith rollout requests that put their tensors in the store, and latents and debug tensors stored as one field per step (#48959, #49693). - Token-space multimodal processing: the MM refactor series, strategy registry, media decoding declaration, per-item feature cache on disk, strategy-owned media loading, Inkling strategy (#46681, #46696, #46699, #46729, #46753, #46767, #46768, #46769, #47355, #47358, #47903, #48285, #48286, #48304, #48305, #48953, #49126).
- LoRA serving: MiMo-V2.6 adapters on native FP8 + MXFP4 checkpoints (#45939), DeepSeek-V4/V4.1 attention and shared-expert adapters (#45231), per-expert MoE LoRA buffers for already-wrapped experts (#46894),
upserton/load_lora_adapter(#47135), LoRA kept off the FlashMLA MegaAttention backend for DeepSeek-V4.1, andkv_b_projadapters kept off the DeepSeek/Kimi MHA cached-prefix paths, with startup rejections for the backends that cannot route cached prefixes through the absorbed MLA path (#49848), which fixes the Kimi-K2.5 native LoRA engine crash described under Model Support & Fixes. - Weight sync: runner weights published over NIXL (#35712), one serialized payload per TP group (#48241), TP node-span probe on the first multimem all-gather (#46682).
- True on-policy: graph logits buffer in the logits dtype, sampling from fp32 probabilities when logits are bf16 (#44517, #45971).
- Models: MiniMax-M3 fused q/k/v + index GEMM writable by online updates (#47612), Muse Spark derived weight caches refreshed after a sync (#49688),
DeepSeekRecipeconvert and encode (#47808), DeepSeek-Spark target verify with forward captures (#48974), MegaGate hash-layer routed experts recorded for R3 (stack commit without a PR), routed-expert caches kept only on the output rank (#49374). - AMD: qwen3 dtype alignment when fused decode omits KV (#47926), import-time device probes skipped without a GPU (#48955).
- Not in the lock: #47960 (
SGLANG_TOPK_V2_MODEexactness modes), needed for the rollout half of DeepSeek-V4.1's smaller-index tie contract.
- Output store: radix transfer plugins, direct reads into caller memory, byte-range reads (sgl-project/sglang#42647, #46288, #47320); for diffusion RL, a tensor stash in
- Megatron-LM:
radixark/Megatron-LM@miles-mainlocked at809217f40d(v0.1.3:2fbe25a442), six commits:moe_permute_accumulation_in_fp32(off by default) removes order-dependent bf16 drift in the unfused MoE permute (radixark/Megatron-LM#132).- Token-dependent absorbed MLA projections, so a LoRA on
linear_kv_up_projreaches the attention output (#133). - TileLang DSA kernel backend ported, restoring
dsa_kernel_backend='tilelang'for GLM-5.2 Bridge LoRA (#134). - MTP loss tracking and activation checkpointing skipped for no-grad forwards, required by Miles' train-mode log-prob passes (#136, #137).
moe_expert_fake_quant="mxfp4"hook for grouped MoE experts (#139); its Miles adapter (radixark/miles#4017) is not in this release, so the hook is inert.
- Megatron-Bridge:
radixark/Megatron-Bridgea0c6d6e09b(was8cd3466d14): canonical MultiLoRA projections, shared-outer expert LoRA, shared-expert and absorbed-MLA multi-LoRA, independent GDNin_projcomponent adapters, TP/SP gradient fixes (radixark/Megatron-Bridge#41, #45, #50, #53–#57) (#3899, #3905, #4542, #4579 @yueming-yuan). - torchtitan: pinned to the v0.4.0 release commit
45e33cd413(was v0.3.0086bf6c166), withspmd_types==0.3.1,torch_remat==0.2.2,torch_checkpointing==0.1.1,grain==0.2.18,attn-gym==0.0.17(tyroandtorchdatadropped); thetorchtitanextra now requirestorch>=2.14(#4537 @Zhichenzzz). - torch_memory_saver:
118b09b3bc(wasb5588e83de): device synchronization before pause backup and unmap, HIP VMM granularity and host access for--offload-train, retained CPU backups, 0.0.10 release (#4247 @guapisolo, #4318 @XinyuJiangCMU). - FlashMLA: the base image's
flash_mlapackage is renamedsgl_flash_mlaand the nv_dev build (withindexer_topk, SM90 and 576-dim sparse prefill for Megatron's CSA forward and the Miles DSA kernels) is installed asflash_mla;sglang-milesimportssgl_flash_mlafirst. - radix_transfer_plugins: installed editable from the
sglang-milescheckout in the CUDA and ROCm images for the output store (#4329, #4478 @guapisolo). - Base image:
lmsysorg/sglang:v0.5.23(was v0.5.22), which also provides thenixlpackage. Prebuilt wheels staycu130-torch214-x86_64/cu130-torch214-aarch64. - Python requirements:
transformers==5.19.0unchanged; addspyrage==1.4.0(age-encrypted Tinker external weights). New extras:torchtitan(torch>=2.14,torchtitan==0.4.0) andnovita(novita-sandbox>=2.0.7).python_requires>=3.11.
Docker
docker pull radixark/miles:v0.2.0- NVIDIA: CUDA 13 only, on
linux/amd64andlinux/arm64, built onlmsysorg/sglang:v0.5.23. The manifest digest issha256:c5e9ebe1dbb02faa094cf5036b57c4f3aadac9a53992f4d143823a6c20fb377c. Sources: miles8e419cc739,sglang-milesfd017154ad, Megatron-LM809217f40d. No CUDA 12 image is built;docker/Dockerfile.cu12remains in the tree with its pins bumped but is deprecated per the docs and not built. - AMD: the ROCm variants build on
lmsysorg/sglang:v0.5.23-rocm724-mi35x,v0.5.23-rocm10-mi35xandv0.5.23-rocm10-mi30xand are published daily torocm/sgl-devby the SGLang nightly; this repository builds them only on manual dispatch;rocm10-mi35xis the CI image androcm724-mi35xis no longer published daily (#4599 @XinyuJiangCMU, #4514 @michaelzhang-ai). - TPU:
python docker/build.py --variant tpubuildsradixark/miles:<tag>-tpu(linux/amd64) fromlmsysorg/sglang:v0.1.0-tpuwith sglang-torchtpu108d775ea4; manual dispatch only, no release tag is produced for it.
Breaking Changes & Upgrade Notes
-
--rollout-max-context-lenand--eval-max-context-lenare removed.- Old:
--rollout-max-context-lencapped/generaterequests;--max-seq-lenexisted only for the agentic wrapper. - New:
--max-seq-lenowns the trajectory cap on both paths. The old flags fail withunrecognized arguments; there is no alias. Under the session server it truncates assembled samples and is passed inmetadata["max_seq_len"]; the agent must stop itself. Setting it also defaults--rollout-max-prompt-lentomax_seq_len - 1. - Action: rename the flag in launch scripts.
- Sources: #4699, #4690 @guapisolo, #3476 @fzyzcjy.
- Old:
-
Prompts are shuffled by default.
-
rollout/raw_rewardchanged meaning; MFU values dropped.- Old:
rollout/raw_rewardwas the mean over the retained batch;rollout/raw_reward_unfilteredheld the pre-filter mean.perf/actor_train_mfuused one dense-attention formula. - New:
rollout/raw_rewardis the mean over every group that reached the dynamic sampling filter; the retained mean isrollout/raw_reward_filtered;raw_reward_unfilteredis gone. MFU is sized per layer from the HF config, so most MoE, linear-attention and sparse-attention models report much lower values (Qwen3.5-35B-A3B forward 47.4 → 16.7 GFLOP/token at 131072 tokens, PR body). - Action: repoint monitoring views and alerts; custom rollouts must publish
rollout/raw_rewardthemselves; do not compare MFU history across the release. - Sources: #4486 @guapisolo, #4096, #4097 @Zhichenzzz.
- Old:
-
Engine weight checksums are no longer implied by
--save.- Old: an event directory (
--save,--dump-details,--save-debug-event-data) turned--log-inference-engine-weight-checksumson. - New: only
--enable-event-analyzer, train fault tolerance or the explicit flag do (FP8 engines dequantize every weight to checksum, which risked OOM). - Action: pass the flag explicitly if you compare checksums.
- Sources: #4319 @yueming-yuan, #4551 @clumsy.
- Old: an event directory (
-
parse_args()returns an immutableAllConfig.- Old: an
argparse.Namespaceany code could mutate or extend;--custom-*-pathfields were strings; hook options from--custom-config-pathYAML orfn.add_argumentssat on the globalargs. - New: a pydantic
AllConfig(fields immutable after parsing; unknown parsed fields raise "Parsed arguments have no configuration owner");args.custom_rm_pathand friends areCustomFunctionConfigobjects with.pathand.config; hook options are declared withfn.config_class = MyConfig(BaseConfig)(legacyadd_argumentsstill adapted) and delivered to the hook as an immutable flat view, deleted from the global config;--custom-config-pathstill works but warns as deprecated, and the example YAMLs are replaced by explicit flags;GenerateState(args)asserts identical args on reuse;reset_arglives inmiles.utils.args.schemaand checks compatibility; worker specs areBaseSpec/BaseCommandSpec/BaseServeSpec;--padded-vocab-sizeis gone from the training entry points (still on the conversion tool);CONVERSION_VERSION5 → 6 re-converts cached torch_dist checkpoints once;getattr/hasattron configs need a# config-access-exemptcomment to pass pre-commit. - Action: mutate in
preprocess_args; readargs.custom_rm_path.path; declare hook options throughconfig_class; useinput.statein generate functions; rename spec imports. - Sources: #3477, #3478, #3522–#3533, #3491, #3497, #3528, #3535, #4361, #4367, #4397, #4415 @fzyzcjy.
- Old: an
-
--lora-typedefaults per backend.- Old: global default
lora; multi-LoRA silently forced--megatron-to-hf-mode bridge. - New: default
None, resolved tocanonical_loraon native raw mode (explicitlorais refused) and forserve_tinker.py,loraon Bridge and torchtitan (torchtitan refuses other values). Multi-LoRA respects the given--megatron-to-hf-mode(bridgeorraw). Tinker checkpoints saved asloraare refused under the canonical default. - Action: drop
--lora-type lorafrom native runs; pass--lora-type lorato a gateway that must restore fused-layout checkpoints; pass the HF mode explicitly. - Sources: #4574, #3905, #4563, #4662 @yueming-yuan.
- Old: global default
-
LoRA checkpoints move to adapter directories.
- Old: per-rank
adapter_megatron_rank*.pt+training_state_rank*.pt, same-layout resume only; torchtitan--savewrote the whole model; Tinkersave_statewrote torch_dist +META.json;--lora-alphadefaulted to 16. - New:
{save}/iter_N/adapter→.adapter.<commit_id>/withadapter_state/,adapter_parametrization.json,training_state.jsonand, when the adapter has a plain-PEFT form, a PEFT export, resumable under any layout. Legacy rank.ptdirs still load under the same layout, and Tinker torch_dist checkpoints load across layouts and slots (Bridge only).--lora-alphadefaults to the source's alpha (required for torchtitan DCP sources).--save/--lora-adapter-pathon adapters require Adam with fp32 masters, DistributedOptimizer and Gloo, and refuse--stream-optimizer-state-to-disk,--indep-dp, SGD and--debug-disable-optimizerwith--save; a failed PEFT export fails the save. Tinker mode refuses--save,--save-intervaland--save-hfand requires--tinker-checkpoint-root; Bridge--save-hfadapters carry thebase_model.model.prefix. - Checkpoints that cannot resume: Bridge canonical Qwen3.5/3.6 adapters saved before the Megatron-Bridge bump, single canonical MoE adapters saved before #4542, torchtitan runs saved under the old whole-model layout (migrate with
--load $OLD --lora-adapter-path $OLD/.../step-N). - Action: copy adapter dirs with symlinks followed (
cp -rL); replace--save Xwith--tinker-checkpoint-root X/tinkerin Tinker mode; namein_proj_b,in_proj_aexplicitly if a Qwen3.5 canonical adapter needs them (they left theattngroup). - Sources: #4608, #4614, #4616–#4619, #4624, #4564, #4579, #4580, #4542 @yueming-yuan.
- Old: per-rank
-
Tinker clients must run SDK 0.36.0.
- Old:
tinker==0.26.2. - New: the gateway answers 400 to any other
sdk_versionatPOST /api/v1/client/config, rejects unsupported sampling and optimizer options instead of dropping them, andrl_loopmust pass a served loss (loss_fn="importance_sampling"or"cispo") instead of itsstable_reinforcedefault. - Action: upgrade to
tinker==0.36.0andtinker-cookbook==0.5.9. - Sources: #4611, #4565 @yueming-yuan.
- Old:
-
torchtitan v0.4.0 on torch 2.14; Python 3.11; pure wheel.
- Old: the image pinned torchtitan v0.3.0 with
spmd_types==0.2.3,tyroandtorchdata, and setup.py had notorchtitanextra;python_requires>=3.10; a platform-taggedcp3X-manylinuxwheel; Megatron's fused cross entropy whenever Megatron was installed. - New: torchtitan v0.4.0, a
torchtitanextra (torch>=2.14,torchtitan==0.4.0) withtorch>=2.14refused below; v0.3.0 DCP checkpoints do not resume (parameter FQNs changed);python_requires>=3.11; onepy3-none-anywheel; log probs use a plain-torch vocab-parallel cross entropy on every backend (last-ULP bf16 gradient differences from the fused kernel, ~9% slower at TP=1). - Action: run on Python 3.11 or 3.12; rebuild the image or
pip install miles[torchtitan]; start torchtitan runs from HF weights. - Sources: #4537 @Zhichenzzz, #4308, #4315 @HJSang.
- Old: the image pinned torchtitan v0.3.0 with
-
LoRA adapters on
kv_b_proj(MLA models: DeepSeek, Kimi, GLM-5.x).- Old: the v0.1.3 engines served cached prefixes for a
kv_b_projadapter through the MHA prefix paths, which applied the adapter delta to the wrong rows (silent corruption on Hopper past 8192 cached tokens, an engine crash on Blackwell). - New: the locked
sglang-milesrejects akv_b_projadapter at startup on thedsa,nsaandaiterprefill backends, with--enable-kv-cache-sharding, and onfa4off SM100; LoRA together with--enable-two-batch-overlapis rejected on every backend. Ontrtllm_mla,fa3,flashinferandflashmlathe chunked prefix cache is disabled forkv_b_projadapters so cached prefixes take the absorbed MLA path; this fixes the Kimi-K2.5 crash on Blackwell and is the correct path for existingkv_b_projLoRA runs; in release CI the GLM-4.7-Flash native LoRA test (fa3) passed with the flag on all engines and its train/rollout log-prob difference was unchanged within noise (0.0329/0.0308 with the flag, 0.0322/0.0312 without). The GLM-5 / 5.1 / 5.2 default LoRA targets no longer includekv_b_proj, and resuming an older GLM-5.x Bridge adapter drops itskv_b_projadapter silently. DeepSeek-V3.2's default targets still includekv_b_projand are rejected at startup. - Action: drop
kv_b_projfrom--lora-target-moduleson DSA-attention models (for DeepSeek-V3.2 until the default is fixed in 0.2.1); do not combine LoRA with--enable-two-batch-overlap; no retuning is expected forkv_b_projLoRA runs that already used the absorbed path. - Sources: sgl-project/sglang#49848 @Rockdu, #4786 @HJSang.
- Old: the v0.1.3 engines served cached prefixes for a
-
Kernel defaults changed on DSA and GDN models.
- Old:
--miles-dsa-topk-backend torch; TileLang DSA backward; eager HF GDN input norm. - New:
--miles-dsa-topk-backend canonical(choicescanonical/torch/flashinfer), cuDNN DSA backward in the tuned rows for sm90/sm100,--linear-attention-input-norm te. Numerics differ at bf16 level (log-prob relative difference 3e-7 to 9e-6, PR body). - Action: pass
--miles-dsa-topk-backend torchand--linear-attention-input-norm hfto reproduce v0.1.3. - Sources: #4001 @Zhichenzzz, #4740 @zianglih.
- Old:
-
DeepSeek-V4 / V4.1 prompts render through SGLang's
DeepSeekRecipe.- New, with no flag: tool JSON drops
strict,developerrenders assystem, a V4 mid-conversation system turn moves first; tool calls withoutid, emptytool_callsand unanswered tool calls (except a trailing turn's) are rejected. Needs the lockedsglang-miles(sgl-project/sglang#47808); an older SGLang imports but cannot render V4/V4.1. - Sources: #4596 @guapisolo. Separately, SGLang v0.5.23 keeps client key order in tool JSON for HF chat templates and DeepSeek-V3.2 (#4548 @Rockdu).
- New, with no flag: tool JSON drops
-
Megatron numerics on resume and log-prob passes.
- Log probs are computed in train mode unless the model has dropout, so training and scoring use the same kernels (step-0
ppo_kl0 on Blackwell) (#4107 @yueming-yuan). The LR scheduler is no longer double-stepped on resume, so learning rates after a resume differ from v0.1.3 (#4530 @clumsy). the train-infer mismatch helper's--tis-level geometricwith--tis-batch-normalizenow runs instead of raising (#4528 @clumsy). THD packing always setstotal_tokens, fixing Mamba numerics and adding one device sync per micro-batch (#4222 @omri-manolo). Qwen3-VL MoE Bridge runs train the whole model again (#4331 @Rockdu).
- Log probs are computed in train mode unless the model has dropout, so training and scoring use the same kernels (step-0
-
Reward and mask semantics.
- gpqa scores only a single committed option; the
boxed_variants ofmath,deepscaler,gemma_mathanddapograde correct answers (every sample scored 0, or -1 for dapo, before);distill_qwenkeeps the full prompt context, trains EOS and raises on a sample without a trainable assistant message; the last kept session turn isTRUNCATEDwhen later turns are dropped, somask_truncated_zero_rewardnow applies;reward_spec.custom_rm_pathnaming a different typed hook, or another function under--group-rmwith a batched hook, raises instead of being ignored. - Sources: #4442, #4443, #4444, #4457, #4490 @Shi-Dong.
- gpqa scores only a single committed option; the
-
GLM stop-token replacement is gated.
--allow-stop-token-replace(requires--use-session-server) defaults to the family default: on for GLM-4.7/5.3 with a warning per replacement, off elsewhere where a replaced stop token now returns HTTP 400. Allowing it together with--sglang-enable-streaming-sessionis rejected at parse time (#4653 @guapisolo). -
Renamed and removed flags.
--fsdp-cpu-offload→--optimizer-cpu-offloadon FSDP and torchtitan (old spelling still parses with a warning) (#3995 @HJSang).- Inkling
--ffn-hidden-sizeis the dense MLP width (16384 for Small, 24576 for the full model), not the expert width;hf_validate_argsasserts on the old value (#4605 @yueming-yuan). --async-submit-within-stalenessis rejected by the rollout with a buffer that lacksadmits(), including multi-policy (#4270 @Arist12, #4733 @HJSang).miles_plugins/models/qwen3_5.pybecame a package;--spec miles_plugins.models.qwen3_5 get_qwen3_5_specstill works, other imports move to.model(#4610 @Rockdu).- The torch profiler traces only
--profile-ranksand records shapes or stacks only with--pytorch-profiler-collect-shapes/-callstack; memory and FLOPs are no longer recorded (#4128 @guapisolo). placement_group.create_training_modelsis removed; drivers usecreate_components(#4724 @HJSang).- NVFP4 exports write +0 codes; reconvert for byte-matching exports or pass
clear_negative_zero=False(#4766 @nanjiangwill). - R3 payloads on hand-built
meta_infoneedprompt_tokensandcompletion_tokens(#4693 @yueming-yuan).
-
Router, ports and run ids.
args.use_miles_routeris replaced byargs.router(--use-miles-routerstays as an alias; config files map); the Miles router registers throughPOST /workers; routing-key headers go only to a router that reads them (#4322, #4321, #4320 @HJSang).- Dynamic service ports end below the kernel's ephemeral range (20000–32767 on a stock kernel); with
MILES_SCRIPT_EXTERNAL_RAY=1add--min-worker-port 0 --max-worker-port 0to your ownray startwhen Ray's default 10002–19999 overlaps it (#4199 @bobboli). - Run ids are
YYMMDD-HHMMSS-xxxx(18 characters, base-36 suffix) (#4472 @HJSang). --miles-router-max-connectionsunset now means no cap; external engines no longer have to sum to--rollout-num-gpus(#4383, #3560 @fzyzcjy).
-
ROCm CI image.
pr-test-rocm.ymldefaults torocm/sgl-dev:miles-rocm10-mi35x; the ROCm 7.2.4 variant is still buildable but no longer published daily (#4514 @michaelzhang-ai). -
Diffusion (radixark/miles_diffusion).
--diffusion-taskis required (t2i,t2v,t2va,ref2va), checked against the family's tasks (#293).--ema-rollout-policy→--rollout-weights {actor,ema}(no alias);--loss-type nftrequires--use-ema --rollout-weights ema;--diffusion-nft-timestep-fraction→--diffusion-nft-timestep-strategy-path+--diffusion-nft-num-dropped-timesteps(default 0;drop_final_stepswith 1 reproduces the old always-drop-the-smallest-sigma behavior); the NFT KL metric iskl_loss; NFT standards were re-recorded (#265, #266, #272, #280, #263).- SFT rows carry
target.visual(and optionalconditions) instead ofmetadata.video; targets must already be on the serving grid; the latent cache version moved to 5 and--seedjoins the key;sft_lossrequires a visual--fsdp-flow-shift; SFT recipes launchtrain_sft.py(#288, #289, #290, #305, #277, #248). --lora-adapter-pathtakes one plain PEFT directory per--update-weight-target-module, in order; rollout-merged LoRA rejects rsLoRA and rank/alpha patterns;--lora-ipc-weight-syncwithout--colocateis rejected (#274, #232).- Custom loss hooks:
LOSS_HOOKS[(loss_type, role)],--custom-*-pathis a parse error outsidecustom_loss, newprepare/ formula /train_outputsignatures with a frozen context, teachers requested throughPreparedBatch.teacher_guidance_scale,create_training_modelsreturns{role: group}, the train-side diffusers scheduler andDiffusersSdeStepBackendare gone, model inputs are{stream: tensor}dicts, step strategies return(sde, ret, stop), sleep/wake helpers live inmiles.backends.fsdp_utils.device_move(#367, #368, #372, #329, #335, #278, #281, #286, #222, #273). - Engine and requests: the image builds the engine from
sglang-miles(pinSGLANG_DIFFUSION_COMMITfor reproducibility);--diffusion-qualitydefaults toexact; adapter-syncing recipes pin--sglang-lora-merge-mode merge; responses carry only the SDE window's latents (--rollout-return-full-trajectoryrestores the old payload);--sglang-scheduler-portis removed;--diffusion-sde-type odewith the Flow-GRPO loss is rejected;--rollout-shuffleis opt-in; trainerNCCL_CUMEM_ENABLE=0only for colocated runs (#257, #315, #216, #222, #358, #320, #238, #314). - Rewards:
--colocate-rewardbecame per-reward flags; mixtures use--custom-rm-argswithweighted_mixture_rmand need--reward-key; new--rm-typevalueshpsv3,openai_api,custom_apineed their config flags (#237, #341, #244). - Numerics that changed on purpose: joint-SFT sigma draws (#303), Krea-2 RMSNorm weights in fp32 (#354), the
isclosesigma_max swap (#279), rollout patches reaching the engine again (#284). - Data plane (opt-in, experimental): custom generate functions call
load_generated_outputsbefore readingsample.generated_output;--use-output-storerejects debug data dumping (#376, #359).
Known Issues
-
DeepSeek-V3.2 default LoRA targets are rejected. Its default targets still include
kv_b_proj, which the lockedsglang-milesrefuses on DSA attention; pass--lora-target-moduleswithoutkv_b_projuntil the default is fixed in 0.2.1. -
Muse Spark Open docs and Muon. The Muse Spark Open page still prescribes
radixark/miles:muse-spark-openand documents--vlmand an HF export tool that are not in this release.--optimizer muonpasses--chunked-muon-master-weights, which the locked Megatron-LM lacks (radixark/Megatron-LM#126 open); use Adam. -
GLM-5.3-Flash torchtitan e2e tests are disabled pending a CI image with GLM-5.3-Flash serving, and the GLM-5.3-Flash pages still point at
radixark/miles:glm53next. -
DeepSeek-V4.1 smaller-index tie contract is trainer-only on this image; the rollout side needs sgl-project/sglang#47960.
-
Docs lag the code in a few places:
docs/advanced/multi-lora.mdstill says the native gateway was not run end to end (a CI test now covers it);docs/developer/ci/02-docker-build.mdstill names the v0.5.22 bases;--router,--p2p-transport,--dynamo-config,--linear-attention-input-normand--use-miles-ext-kernelshave no CLI-reference rows; the geo3k multi-turn page still documentsgeo3k_vlm_multi_turn_config.yaml, which no launcher reads any more. -
Foundations that are not features yet: the Dynamo router (parts 1–4 of 13;
--router dynamois rejected), themiles-ext-kernelssubmodule (scaffold, not cloned or installed by the image), and the MXFP4 fake-QAT hook in Megatron-LM (Miles side #4017 open). -
Evidence limits: TPU runs are single-host smoke and bench recipes; NIXL is validated on one node only; native LoRA in disaggregated and PP>1 modes, Bridge single-LoRA PP>1 save/load, legacy
.ptadapter loads and shared-outer Tinker slot checkpoints in the new format are validated by hand or not on GPU; the MiniMax-M3.1-Flash full recipe has not been executed; AMD changes are not covered by release CI. -
Diffusion (radixark/miles_diffusion): DMD2 has no e2e yet and its recipe has not run at the 16-GPU scale;
setup.pysays 0.2.1 while no release has been cut; the README lacks Krea-2 and FLUX 3 rows; the diffusion docs in this release predate the data plane, DMD2, score centering and HPSv3.
All Contributors
@amitw-nv @AndyDai-nv @Arist12 @bobboli @CeerDecy @clumsy @fzyzcjy @guapisolo @HJSang @michaelzhang-ai @nanjiangwill @omri-manolo @Rockdu @Shi-Dong @Weyaxi @XinyuJiangCMU @yueming-yuan @yushengsu-thu @Zhichenzzz @zianglih
20 contributors, counted as authors of the 400 PRs between v0.1.3 and v0.2.0; bot commits and AI co-author trailers are excluded. Commit trailers also credit Zhiyao Jiang, bingxche and Tao Lin as co-authors.
Full Changelog: v0.1.3...v0.2.0