Project Chronos (Experimental)

A storage-aware MoE stack built for SSD+DRAM hybrid inference, with a full six-stage training pipeline.

中文文档

Chronos forward stack. The embedding lookup feeds a prelude of $P$ blocks, a recurrent region of $G$ blocks run $R$ times with conditioning, and a coda of $C$ blocks, then a tied LM head. Prefetch, lookahead and expert residency hang off the recurrent region.

---

Documentation

Start here: docs/README.md is the full index with reading routes and figures.

Area	Chapter	What it answers
Entry point	OVERVIEW.md	How do computation, routing and residency connect?
Entry point	KERNEL_ARCHITECTURE.md	What is the implementation-exact end-to-end graph?
Reference	SYMBOLS.md	Axes, units and configuration fields.
Reference	THEORY.md	Stability, information and resource derivations with assumptions.
Model	ATTENTION.md	MLA/SWA projection, caching, reconstruction and masking.
Model	MOE_ROUTING.md	Dispatch, shared averaging, fallback and regularizers.
Model	CLUSTER_LOCALITY.md	Fixed-partition mutual-information objective.
Model	LOOKAHEAD.md	Future-token expert-distribution prediction and training.
Model	LOOP_DEPTH.md	Matched-step and matched-time recurrent-depth experiments.
Runtime	LAZY_EXPERTS.md	Expert-page loading, admission and materialization.
Runtime	RESIDENCY.md	Eviction policies and residency counters.
Runtime	SPEC_DECODE.md	Draft/verify sampling correctness and cost.
Runtime	LONG_CONTEXT.md	Chunked loss, blockwise attention and recomputation.
Evidence	REPRODUCIBILITY.md	Evidence and controls required for a claim.
Serving	vllm_integration.md	Serving interfaces and integration scaffold.
中文	READING_GUIDE_ZH.md	中文阅读路线、核心机制与符号导引。

---

Why reactive MoE offload breaks down

Mainstream MoE models such as Mixtral, DeepSeek-MoE, and Qwen-MoE make routing decisions at decode time, token by token. The model discovers which experts it needs only when the next token is already on the critical path. If those experts are not resident in VRAM, generation stalls while weights are moved from SSD to RAM to VRAM.

That is not a tuning issue. It is an architectural mismatch.

Most offload runtimes assume the model was originally designed for full VRAM residency, then try to patch storage pressure afterward. On consumer hardware, that usually means the decode loop ends up paying the IO bill over and over again.

---

Chronos in one line

Move IO into prefill. Move synchronization down to the expert-event level.

sequenceDiagram
    autonumber
    participant Prompt as Prompt tokens
    participant Intent as IntentClassifier
    participant SSD as Clustered SSD cache
    participant RAM as Pinned RAM
    participant H2D as H2D stream
    participant GPU as Decode compute stream
    participant LR as LookaheadRouter

    rect rgb(236, 248, 255)
        Prompt->>Intent: prefill reads the whole prompt
        Intent->>SSD: predict generation-level expert set
        SSD->>RAM: mmap cluster-packed .ctsr files
        RAM->>H2D: enqueue hot experts before first decode token
        H2D-->>GPU: per-expert ready events
    end

    loop every decode token
        GPU->>LR: block-0 hidden state
        LR->>H2D: prefetch experts for t+1..t+K
        GPU->>GPU: forward token t with resident experts
        GPU->>H2D: wait only on the expert events it needs
    end

Loading

Two things matter here:

1. Prefill-time loading: PrefillScheduler and IntentClassifier bulk-load predicted experts before the first decode token.
2. Per-expert event sync (M3): promote_to_vram(blocking=False) records a torch.cuda.Event on the _h2d_stream, and the compute stream waits only on the experts it actually needs. No more stream.synchronize() for the whole system.

Under a simulated 30 ms SSD delay, the new path keeps 35 ms+ per token of pipeline slack compared with the older blocking path.

---

v3 kernel: recurrent-depth MoE (0.3.0)

v3 reuses a few physical layers many times. A prelude runs once, each shared recurrent group loops recurrent_depth times with a contractive update h ← h + dt·(block(h + e_loop) − h), and a coda runs once. The 50M demo has 6 physical layers that execute as 18 virtual layers (prelude 1, 4 recurrent layers × depth 4, coda 1).

flowchart LR
    X["tokens"] --> P["Prelude<br/>runs once"]
    P --> G["Recurrent group<br/>MLA / SWA + top-2 MoE"]
    G -->|"loop d = 1..D<br/>contractive step, ACT halt"| G
    G --> C["Coda<br/>runs once"]
    C --> H["LM head"]
    P -. "lookahead source" .-> LR["LookaheadRouter<br/>expert prefetch"]

Loading

What changed from v2

	v2	v3
Attention inside the loop	Iteration 0 only (first_only)	Every iteration
KV cache per recurrent layer	1 slot; the every-iteration attention switch rewrote it each iteration (measured prefill+decode vs full-forward gap 0.94)	shared_first: Q recomputed each iteration, K/V from iteration 0, so KV slots = physical layers
Routing	top-1 (gate received no CE gradient)	top-2 + 1 shared expert
MoE dispatch	Python loop over experts with .item() syncs	Sorted grouped dispatch: gather_mm on MLX, argsort + index_add_ on Torch
Early exit	Host sync every step	Tensor-only ACT, exact and batch-invariant; decode can exit per token without KV gaps
Decoding	Plain autoregressive	Recurrent-depth self-speculative decoding

per_iteration (one KV slot per virtual layer) is kept as an ablation. Old v2 checkpoints load as first_only; their logits match a stored golden tensor within 1e-5.

Self-speculative decoding (no extra parameters). The same model drafts γ tokens at a shallow depth and verifies them with one full-depth batched forward pass. Greedy output is token-for-token identical to full-depth greedy. Sampling uses rejection sampling, which preserves the full-depth distribution. After a rejection, the cache is truncated: MLA slices c_kv, and the sliding-window cache keeps window + γ_max entries. Torch (chronos/runtime/speculative.py) and MLX (chronos/mlx/speculative.py, ChronosMLXInferenceEngine.generate(speculative=True)) share the algorithm. Tests cover greedy identity, the rejection distribution and rollback against a fresh prefill on both backends. The Web UI Inference tab has a switch for it, with draft depth and γ.

For the 50M demo checkpoint it is exact but not faster on any runtime we measured (pretrain step-13,996 checkpoint of the demo in demos/translation, greedy, batch 1, the demo's dev prompts, draft depth 1, γ 4; exact conditions in demos/translation/README.md; reports in reports/speculative/):

Runtime	Plain tok/s	Speculative tok/s	Ratio	Acceptance
MLX GPU bf16, all experts in DRAM	112.0	107.4	0.96×	42 %
MLX GPU bf16, lazy experts (12 of 216 pages resident)	43.7	32.0	0.73×	42 %
Torch MPS fp32, all experts	32.1	25.9	0.81×	40 %
Torch MPS fp32, lazy experts	17.0	14.7	0.86×	40 %

At this size a full-depth step costs about as much as a batched verify pass. With lazy experts, drafts and verify passes also route to different experts, so speculative decoding read 1.4–1.6× more expert pages from SSD. Greedy outputs were identical in CPU fp32 and Torch MPS fp32. In GPU bf16, near-tie tokens flipped in 2 of 24 sentences. The switch therefore stays off by default.

Precision chain. One compute dtype runs end to end. FP32 is used only where it pays: RMSNorm statistics, SDPA softmax, RoPE tables, router softmax, LM-head logits and optimizer master weights. The MLX↔Torch bf16 bridge is bit-exact. auto picks bf16 on MLX GPU and CUDA, and fp32 on CPU and MPS (MPS bf16 measured slower; see below).

Tooling added in v3

Command	What it does
chronos tokenizer	Self-built byte-fallback BPE with script-aware pre-tokenization and language-balanced merges; writes an HF-compatible tokenizer.json + chat template
chronos longrun --spec recipe.json (alias run)	Time-budgeted MLX run: warmup + cosine per phase wall-clock budget, NaN rollback, bit-exact resume, validation CE; tasks add their own metrics by subclassing LongRun (task_eval)
chronos serve	Ollama/OpenAI protocol server on the native runtime (see below)
chronos monitor	Read-only training monitor page + JSON API (see below)

chronos.model.factory.fit_config searches for a configuration that matches total and active parameter targets.

Kernel and extensions. The chronos package trains general models and carries no task-specific code. Task tooling lives in extensions that import the kernel. The en↔zh translation demo is one: demos/translation (chronos-translate: corpus pipeline, BLEU/chrF runs, model bundles; see its README). It is not part of the Project_Chronos wheel and is not published on PyPI; install it from a checkout with pip install -e demos/translation. The 50M en→zh status, the released model chronos-enzh-50m-sft (files in demos/translation/models/) with its measurements, and the model bundle are documented in demos/translation/README.md.

Inference and memory options (all off by default). Each option is a config field and a chronos serve / Inference-tab override. Tests compare each one with its reference path on Torch and MLX.

Field	Values (default first)	Effect
kv_cache_dtype	auto, int8	int8 stores MLA c_kv and SWA k/v per token in groups of 32 with a scale; rollback and speculative decoding work on it
mla_decode	expanded, absorbed, auto	absorbed scores against c_kv without expanding K/V. auto absorbs only short cached chunks; it is opt-in because a bf16 prefill (expanded) followed by absorbed decode drifted from the full forward pass (max logit gap 0.06, measured)
mla_cache_rope	false, true	Also caches the roped key, trading memory for decode speed
router_scope	per_iteration, shared	shared reuses iteration 0's experts for the rest of the loop, so fewer experts are resident per token
draft_exit_kl_coef, draft_exit_depth	0, 1	Training term: KL from the shallow exit to the detached full-depth logits, which raises speculative acceptance
max_resident_experts, ram_budget_mb	unset	Hard caps on the fast tier and the RAM tier. LRU eviction never evicts experts the lookahead router predicts. Prefetch covers every loop iteration of a virtual layer

Web UI: the Tokenizer tab (extract → train → evaluate) and the Serve tab (start/stop chronos serve; the API key goes through the environment, never argv).

---

Native kernel (0.3.2): minimind removed

Up to 0.3.1a3, Chronos took its config base, RoPE tables, gated FFN, dense attention and trainer utilities from minimind. It loaded them at import time: chronos.deps git-cloned the master branch into ~/.cache/chronos/minimind-master/ and put that folder first on sys.path. 0.3.2 replaces those pieces with chronos/model/kernel.py. Nothing is downloaded, and sys.path is not changed.

Why it was migrated

• Reproducibility. The clone followed an unpinned master, so two installs could run different kernel code.
• Offline installs. On a machine without the cache, import chronos ran git clone and failed without network access.
• Namespace collisions. minimind's top-level model, trainer and dataset packages went first on sys.path. They shadowed any user module with the same name.
• Dense mode was broken. use_hybrid_attention=False used minimind's Attention, which crashed in every recurrent mode (The size of tensor a (8) must match the size of tensor b (1024)). It also could not use the shared KV cache.
• BF16 checkpoints were overwritten as FP16. The Torch pretrain save called minimind's lm_checkpoint. That function rewrote the release file chronos_<H>_moe.pth as FP16 right after Chronos had saved it in BF16.
• Trainer logs were lost. The Web UI captured trainer output by patching trainer_utils.Logger. The trainers import Logger by name at load time, so the patch never reached them.

What the native kernel does

• ChronosBaseConfig keeps the old field names and defaults. RoPE/YaRN tables and FFN parameter names are unchanged too, so existing checkpoints and sidecars load as before.
• One FP32 RoPE table per (dim, base, scaling, device) is shared by all layers and grows on demand. Before, every attention layer kept its own copy.
• GlobalAttention is the dense GQA path. It is the sliding-window kernel without a window, so all three recurrent modes, int8 KV, padding masks and speculative rollback work on it.
• Logging uses a sink registry (add_log_sink). Resume state goes to chronos_<H>_moe_resume.pth, and the release checkpoint keeps its dtype.
• The 6400-token tokenizer that old checkpoints use ships as data only (chronos/assets/tokenizers/legacy_6400, Apache-2.0, sha256 recorded in its NOTICE).

Old vs new, measured 2026-11-08

Setup: Apple M5 Max, PyTorch 2.10 on CPU, FP32, 4 threads, median of 5–7 runs, each kernel in its own process, runs alternated twice. A 50M training run was using the GPU at the same time, so absolute timings carry noise. Old = 0.3.1a3 + minimind 24896cd.

Check	0.3.1a3 + minimind	0.3.2 native
Logits, out/chronos_256_moe.pth and the 50M demo checkpoint (full, padded and prefill+decode paths)	reference	bit-identical (max abs diff 0)
Random-init state_dict and config values (same seed)	reference	identical
import chronos with no cache and no network	fails (git clone error)	works
minimind on sys.path, top-level model module imported	yes	no
RoPE table memory, 50M shape (1024 positions)	3.50 MiB	0.50 MiB
RoPE table memory, 768H base shape (32768 positions)	216 MiB	24 MiB
Dense GQA mode (use_hybrid_attention=False)	crashes	works; prefill+decode vs full forward ≤ 1e-6
Release checkpoint dtype after BF16 Torch pretrain	FP16	BF16
Trainer lines captured by the Web UI sink	0 of 2	2 of 2
Train step, 50M shape, batch 4×128	0.318–0.326 s	0.317–0.321 s
Prefill 128 tokens, 50M shape	40.0–40.7 ms	41.4–42.8 ms
Decode tokens/s, 50M shape	146–157	146–151

A repeat with 40 prefill runs per process gave medians of 64–86 ms (old) and 74–86 ms (new) as the training load changed; neither kernel was consistently faster. Compute speed is unchanged within noise: the kernel keeps the same math. The gains are in memory, robustness and the bugs listed above. The parity check runs in CI as tests/test_native_kernel.py, which covers config defaults, RoPE bit-identity, dense mode, imports, logging, resume and the bundled tokenizer.

chronos.deps remains as a shim: get_tokenizer_path() returns the bundled tokenizer (CHRONOS_DEFAULT_TOKENIZER overrides it), and ensure_minimind() is a no-op with a deprecation warning. The comparison tools in tools/compare_minimind_chronos*.py still use minimind as an external baseline, from a checkout you provide (--minimind-dir or CHRONOS_MINIMIND_DIR). benchmark_compare.py now compares Chronos with a dense GQA baseline built from the same kernel.

MLX training: compiled step and activation recomputation (0.3.2)

Up to 0.3.1a3 the MLX trainers ran every micro step eagerly and kept every activation for the backward pass. 0.3.2 adds both missing pieces to chronos longrun and to the Stage 1–2 MLX trainer (pretrain, SFT):

• Compiled micro step (chronos/mlx/training/compiled.py, on by default). Loss and gradients run as one mx.compile graph per signature. The executed recurrent depth, the ACT switch, input shapes and plain config values are part of the signature, so a step never reuses a stale graph. Weights and the RNG state are compile inputs and outputs, so dropout and the cache-miss simulation consume the RNG stream exactly as before. Multimodal batches, lazy (paged) experts and any signature that fails to trace run eagerly, and the reason is logged once. CHRONOS_MLX_COMPILE=auto|on|off or the recipe's extra.mlx_compile switches it.
• Activation recomputation (chronos/mlx/checkpointing.py, activation_checkpointing=true; the planner's memory-constrained tiers turn it on). Each block call becomes one recomputed region. A plain mx.checkpoint is wrong here in two ways, both measured on mlx 0.32.4. Dropout masks and the cache-miss draws would be drawn again during recomputation; they are recorded in the forward pass and replayed. A region that returns the same graph node twice keeps the gradient of only one copy. A block's KV cache and its shared view are such a pair whenever the SWA window covers the sequence (a full-length slice is a new Python object over the same node), and the gradients were 37–40% off. Region outputs are therefore packed into one array per dtype.

Measured 2026-11-08. Setup: Apple M5 Max, mlx 0.32.4, bf16, the 50M demo recipe's model config (demos/translation/recipes/) from random init, the demo's real training shards, the recipe's step sizes (SFT 8 micro-batches per step, pretrain 16). Median of the second half of 16 (SFT) or 12 (pretrain) steps. The 72-hour run was training on the same GPU, so absolute timings carry noise.

Micro step	SFT s/step	SFT peak memory	Pretrain s/step	Pretrain peak memory
Eager (0.3.1a3)	2.915	5.62 GB	7.154	6.07 GB
Compiled (new default)	2.700 (1.08×)	5.86 GB	6.057 (1.18×)	6.28 GB
Recomputation + compiled	3.148	3.50 GB	7.357	3.62 GB
Recomputation, eager	3.663	3.42 GB	—	—

The loss curves of all variants follow the eager one: within 2e-3 over 16 SFT steps and 2e-4 over 12 pretrain steps. The first compiled step also traces the graph. Recomputation cuts peak memory by 38–40%. With compilation on, a step then takes 17% (SFT) or 21% (pretrain) longer than the compiled step without recomputation, and 8% or 3% longer than the old eager step. At 50M parameters memory is not the limit, so recomputation stays off by default; it is meant for larger models or longer sequences on the same machine.

Exactness is tested per micro step on the same weights, batch and RNG state inside the real LongRun loop (tests/test_longrun.py, tests/test_mlx_checkpointing.py). On CPU FP32 the recomputed step's gradients match the eager ones within 4e-7 relative, the compiled step's within 1.7e-5, and with both on within 1e-4. That is FP32 reassociation: scaling the weights by (1 + 1e-7·N(0,1)) moves the eager gradient by 1.8e-4. On the M5 GPU, FP32 matrix products round like TF32. There, the eager gradient itself differs from the CPU FP32 one by 1.5e-2, and the recomputed one differs from the eager one by 1.4e-4.

---

Three-tier storage (M1: cluster-aware safetensors)

flowchart TB
    subgraph GPU["VRAM / Metal unified memory"]
        Dense["Dense layers"]
        Shared["Shared experts<br/>always resident"]
        Hot["Predicted hot experts<br/>promoted before use"]
        Events["Per-expert CUDA events<br/>no global stream sync"]
    end

    subgraph RAM["Pinned RAM staging tier"]
        Buffer["Prefetched expert tensors"]
        LRU["RAM LRU cache"]
    end

    subgraph Disk["NVMe SSD"]
        C1["cluster_000.ctsr"]
        C2["cluster_001.ctsr"]
        Manifest["cluster_manifest.json<br/>expert -> cluster map"]
    end

    Manifest --> C1
    Manifest --> C2
    C1 -->|"sequential mmap read"| Buffer
    C2 -->|"sequential mmap read"| Buffer
    Buffer --> LRU
    LRU -->|"non-blocking H2D"| Hot
    Hot --> Events
    Shared --> Events
    Dense --> Events

Loading

cluster_manifest.json and .ctsr files are produced offline by Louvain clustering. At runtime, Chronos uses safetensors.safe_open(...).mmap to bring an entire expert cluster into RAM in one shot, turning random reads into mostly sequential reads.

Cache misses degrade gracefully instead of stalling

Even the worst case does not hard-stop generation:

# Pure tensor math: no Python branch, no graph break under torch.compile
output = avail[i] * expert_output + (1.0 - avail[i]) * shared_expert_output

The shared expert is always resident, so generation continues while the missing expert finishes loading in the background. For exact lazy/offload comparison modes, Chronos synchronously materializes only the selected missing expert and evicts low-LRU experts to stay inside the resident budget; it does not silently full-load all experts. Quality degrades smoothly only when fallback mode is explicitly enabled.

---

Dual-layer routing with supervised lookahead (M2)

flowchart LR
    Prompt["Prompt / prefill context"] --> IC["IntentClassifier<br/>generation-level expert prior"]
    IC --> Budget["ExpertPredictor<br/>budgeted hot set"]
    Budget --> Prefill["PrefillScheduler<br/>bulk preload before decode"]

    Token["Decode token t"] --> Block0["Transformer block 0"]
    Block0 --> LR["LookaheadRouter<br/>Q_t^(1..K)"]
    LR --> Future["Future expert predictions<br/>t+1 ... t+K"]
    Future --> Async["AsyncPrefetcher<br/>prefetch queue"]

    Prefill --> Cache["CacheManager / ExpertStore"]
    Async --> Cache
    Cache --> MoE["ChronosMOEFeedForward<br/>resident expert or shared fallback"]

Loading

	IntentClassifier (Layer 1)	LookaheadRouter (Layer 2)
When it runs	Once during prefill	Every token during decode
Input	Full prompt (up to 512 tokens)	Hidden state after Block 0
Output	Expert set for the full generation	Expert IDs for t+1, t+2
Training target	Supervised from real activation logs	L_lookahead supervised by real router decisions at t+k
Parameter count	~10-15M	~2M

Before M2, the lookahead head was just a head with no real supervision. M2 adds a proper soft-target objective:

$$L_{\mathrm{lookahead}} = \frac{1}{|K_{\mathrm{valid}}|} \sum_{k \in K_{\mathrm{valid}}} \mathbb{E}_{b,t} \left[ - \sum_{e=1}^{E} \mathrm{sg}(P_{b,t+k,e}) \log Q_{b,t,e}^{(k)} \right]$$

That turns the lookahead router into an actual predictor of future routing, instead of a best-effort heuristic.

---

Full training stack: Stage 1 -> Stage 6

Each stage has its own entry script, and every one of them inherits the shared Chronos loss mixer: balance loss, temporal locality loss, lookahead loss, and, for alignment stages, a router KL anchor that keeps RL/DPO updates from destroying cache locality.

flowchart LR
    P1["Stage 1<br/>Pretrain<br/>CE + Chronos mix"] --> P2["Stage 2<br/>SFT<br/>assistant-token CE + mix"]
    P2 --> P3["Stage 3<br/>DPO<br/>preference loss + anchor"]
    P2 --> P4["Stage 4<br/>ORPO<br/>NLL + odds-ratio + anchor"]
    P4 --> P5["Stage 5<br/>GRPO<br/>rollout reward + KL + anchor"]
    P5 --> P6["Stage 6<br/>Distill<br/>KD + CE + anchor"]

    subgraph Shared["Shared Chronos training terms"]
        B["raw load-balance aux"]
        T["temporal locality"]
        LA["supervised lookahead"]
        A["router KL anchor<br/>alignment stages"]
    end

    Shared -. applied to .-> P1
    Shared -. applied to .-> P2
    Shared -. applied to .-> P3
    Shared -. applied to .-> P4
    Shared -. applied to .-> P5
    Shared -. applied to .-> P6

Loading

Stage	Script	Core objective	Router KL anchor (default lambda)
1 Pretrain	train_chronos.py	CE + balance + temporal + lookahead	0.0 (off)
2 SFT	train_chronos_sft.py	SFT loss + shared Chronos mix	0.01 (weak)
3 DPO	train_chronos_dpo.py	DPO log-sigma(beta * logits) + mix	0.10 (strong)
4 ORPO	train_chronos_orpo.py	NLL + lambda * ORPO term	0.10
5 GRPO	train_chronos_grpo.py	PG * A - beta * KL with ToyReward or pluggable LMRewardModel	0.10
6 Distill	train_chronos_distill.py	`alpha * T^2 * KL(student

Training dtype and resource policy:

• --dtype auto is the default. One compute dtype covers weights, activations, KV cache and expert pages; FP32 is kept only inside fused kernels (SDPA softmax, RMSNorm statistics), for RoPE tables, router softmax, loss logits and the optimizer's FP32 masters/moments. auto resolves to BF16 on MLX GPU, BF16 on CUDA (FP16 if unsupported, with loss scaling), and FP32 on CPU and MPS. On MPS, native BF16 measured slower than FP32 (Chronos 50M on an M5 Max with torch 2.10: 0.749 vs 0.619 s per training step, 49.1 vs 37.3 ms per decoded token). Pass --dtype bfloat16 explicitly to trade speed for half the weight and KV memory.
• CPU training configures PyTorch to use physical cores by default. Override with --cpu_threads or --cpu_budget_percent.
• On macOS, MPS/MLX training forces DataLoader workers to 0 by default to avoid Metal command-buffer crashes from multiprocessing. CPU/CUDA still use worker processes; advanced users can override the guard with CHRONOS_ALLOW_METAL_DATALOADER_WORKERS=1.
• Native MLX training pushes UI logs, scalar readouts, and chart points every log_interval steps, and Web UI Stop is checked at each batch boundary.
• The Web UI writes a warning-only <checkpoint>.verify.json after each stage. It checks no-mask vs all-available MoE parity and, on Apple Silicon, MLX prefill logits against the PyTorch CPU baseline.

Native MLX training is a separate Apple Silicon backend, not torch.to("mlx"). The six-stage MLX trainer mirrors the PyTorch loss stack in chronos.mlx.*: masked CE, DPO/ORPO/GRPO/distillation losses, load balance, temporal locality, lookahead soft-target supervision, lookahead top-k hit loss, and router KL anchor. Numerically sensitive pieces run in FP32 even when model weights use BF16/FP16, which is why auto prefers BF16 over FP16 on MLX GPU: FP16 has too little exponent range for router softmax, CE, and Adam moments on small unstable training runs.

The MLX Web UI trainer reports the same fields as CPU/MPS training: step/loss/steps-per-second/ETA, checkpoint-save events, stop events, and verify results. Stop is cooperative at batch boundaries and save intervals write a valid .pth plus sibling .config.json, so MLX stages can resume or feed the PyTorch/export pipeline.

Memory, media and tools on MLX:

• Every MLX model caps the MLX buffer cache at min(RAM/8, 16 GiB) before it allocates weights (chronos.mlx.memory). A recipe's extra.mlx_cache_limit_gb or CHRONOS_MLX_CACHE_LIMIT_GB may lower the cap, never raise it. The cap bounds reclaimable cache only; weights, activations and gradients are not limited.
• Image and audio encoders and the gated cross-attention fusion run on MLX in all six stages, with the same weights and numerics as PyTorch (chronos.mlx.multimodal). In DPO and ORPO, both answers of a pair condition on the prompt's media. In GRPO, every rollout conditions on it. Preference and GRPO records carry media in the same image, audio and *_embeds fields as SFT. The PyTorch DPO, ORPO and GRPO trainers are text-only, so auto routes media data in those stages to MLX.
• Tool calls:
  • Training: the chat template renders tools, assistant tool_calls and tool results. chronos.data.chat_records.encode_chat gives each token a loss flag (assistant text and calls are learned; everything else is context), and write_sft_records writes .lmask shards that SFTRecords reads.
  • Serving: when a request carries tools, chronos serve parses the model's <tool_call> output into Ollama message.tool_calls or OpenAI tool_calls, in streaming and non-streaming responses.

The full six-stage comparison harness lives in tools/compare_minimind_chronos_v3.py (needs a minimind checkout via --minimind-dir).

---

Backend dispatch (M5)

flowchart TD
    Request["User request<br/>train / inference / WebUI / CLI"] --> Dispatcher["BackendDispatcher"]
    Dispatcher --> Probe["Capability probes"]
    Probe --> CUDA["CUDA<br/>training + inference"]
    Probe --> MPS["MPS<br/>training + inference"]
    Probe --> MLX["MLX<br/>Apple Silicon native path"]
    Probe --> CPU["CPU<br/>portable fallback"]
    Probe --> EXT["Extension hooks<br/>Vulkan / OpenCL"]
    Dispatcher --> Choice["Priority + availability decision"]
    Choice --> Runtime["Chronos runtime / trainer"]

Loading

from chronos.backend import BackendDispatcher

d = BackendDispatcher()
d.available()   # ['mlx', 'mps', 'cpu'] on Apple Silicon
                # ['cuda', 'cpu']        on NVIDIA hosts
d.select()      # choose the best available backend automatically
d.describe()    # human-readable capability summary

• First-class backends for training and inference: cpu, mps, cuda, mlx
• Inference-only / experimental: vulkan when PyTorch was custom-built with USE_VULKAN=ON
• Third-party extension hook: opencl, via chronos/backend/ext/opencl.py:PROBE()
• Apple Silicon policy: inference auto still prefers MLX; training keeps MLX on the native chronos.mlx.* path instead of calling torch.model.to("mlx").

Honest note: upstream PyTorch does not ship a real OpenCL backend, and Vulkan support is still niche. Chronos provides a dispatcher seam so external integrations can plug in cleanly without touching core code.

MLX lazy/offload runtime

MLX uses Apple unified memory, so Chronos treats "VRAM" and "RAM" as logical tiers rather than physical buses. The native lazy runtime still enforces the same offload contract as CUDA/MPS/CPU:

• Hot slots hold only the execution-budget experts materialized as live MLX modules.
• Warm cache holds a bounded prediction buffer loaded from .ctsr safetensors, never a hidden copy of every expert.
• Cold storage is the checkpoint/export reader or per-expert cluster cache. In lazy mode, live experts are replaced by placeholders after cache creation; Chronos does not keep a _saved_live full-expert repair cache.
• Lookahead predictions are queued into the warm cache and only promoted when ready. A true miss synchronously materializes the selected expert only; it does not full-load the model.
• MLX attention dynamically grows RoPE lookup tables past max_position_embeddings, so long prompts plus decode do not fail at token 257/513/etc.

Inference compare/sweep reports real hot/warm counts, resident hit rate, prediction hit rate, sync SSD loads, MLX active/cache/peak memory, process RSS, prefill/decode time, and tokens/sec. A lazy run is not considered offload-ready unless deterministic exact-lazy output matches full-DRAM output and fallback weight stays zero.

---

Hugging Face and vLLM compatibility (M5)

• ChronosForCausalLM subclasses PreTrainedModel and registers AutoConfig and AutoModelForCausalLM, so loading does not require trust_remote_code:

  from transformers import AutoModelForCausalLM
  
  model = AutoModelForCausalLM.from_pretrained("./out_dir")

• chronos.model.hf_io.save_chronos_pretrained and load_chronos_pretrained emit standard model.safetensors + config.json, while also carrying cluster_manifest.json and .ctsr files for expert-cache layout. Roundtrip logit drift is 0.00e+00.

• chronos.serving.register_chronos_with_vllm() registers Chronos with the vLLM ModelRegistry when vLLM is installed. If vLLM is absent, it prints an install hint and exits cleanly. Worker-side mask injection is documented in docs/vllm_integration.md.

---

Compared with existing offload stacks

Feature	llama.cpp offload	vLLM offload	Project Chronos
Expert prediction	None	None	Predictive (IntentCLF + LookaheadRouter)
Lookahead training	n/a	n/a	Supervised L_lookahead (M2)
IO timing	During decode, blocking	During decode, blocking	During prefill, async
Decode pipeline	Synchronous	Synchronous	Dual-stream + per-expert events (M3)
Cache miss behavior	Hard stall	Hard stall	Soft gating, zero hard stall
Disk format	GGUF	safetensors	Cluster-packed safetensors (.ctsr)
Training integration	Post-hoc patch	Post-hoc patch	Native six-stage stack + router KL anchor
Backend dispatch	Compile-time fixed	CUDA only	cpu / mps / cuda / mlx + extension hooks
Apple Silicon support	Partial	No	Full MLX backend
Hugging Face compatibility	GGUF only	Yes	Yes, with expert-cache metadata
vLLM compatibility	n/a	Native	Optional adapter

---

Objective

$$L_{\mathrm{total}} = L_{\mathrm{base}} + \lambda_{\mathrm{bal}} L_{\mathrm{aux}} + \lambda_{\mathrm{tmp}} L_{\mathrm{temporal}} + \lambda_{\mathrm{LA}} L_{\mathrm{lookahead}} + \lambda_{\mathrm{anc}} L_{\mathrm{routerKL}}$$ $$L_{\mathrm{aux}} = E \sum_{e=1}^{E} load_e \cdot \overline{p}_e$$ $$L_{\mathrm{temporal}} = \mathbb{E}_{b,t} \left[ \left\| P_{b,t,:} - P_{b,t-1,:} \right\|_2^2 \right]$$ $$L_{\mathrm{routerKL}} = D_{\mathrm{KL}} \left( \pi_{\theta}^{\mathrm{router}} \| \pi_{\mathrm{ref}}^{\mathrm{router}} \right)$$

• L_base: stage-specific objective (CE, DPO, ORPO, GRPO, or distillation).
• L_aux: the unscaled MoE load-balance auxiliary term; Chronos applies lambda_bal once in chronos_loss_term.
• L_temporal: encourages adjacent tokens to reuse similar expert distributions.
• L_lookahead: soft-target cross entropy from the real future router distribution to the lookahead prediction. sg(...) means stop-gradient.
• L_routerKL: keeps alignment-stage updates from destroying the routing layout captured at stage start.

All lambda terms are searchable with Optuna TPE, together with structural hyperparameters such as hidden_size, num_experts, and kv_latent_dim.

---

Installation

pip install Project_Chronos

Or from source:

git clone https://github.com/FonaTech/Project_Chronos
cd Project_Chronos
pip install -e ".[dev]"

MLX (Apple Silicon):

pip install "Project_Chronos[mlx]"

Optional extras (versions pinned):

pip install "Project_Chronos[serve]"      # chronos serve / chronos monitor (fastapi, uvicorn)
pip install -e demos/translation          # translation demo extension (chronos-translate), from a checkout only, not on PyPI

vLLM serving (optional, Linux + CUDA only):

pip install vllm

No external kernel: since 0.3.2 the kernel is self-contained (chronos/model/kernel.py). Installing and importing Chronos needs no git and no network. See Native kernel (0.3.2) and THIRD_PARTY_NOTICES.md.

Requirements: Python 3.10+, PyTorch 2.6+

---

Quick start

Web UI (11 pages, 4 languages)

chronos-ui                       # or: python chronos_app.py
chronos-ui --lan                 # also open it to the local network (off by default)
chronos-ui --legacy              # the previous Gradio console

A FastAPI server with a single-page app (chronos/webui/). It listens on 127.0.0.1 without a login. Pages: Model (v3 presets, the auto planner below, a live parameter / memory estimate), Data (scan one or several datasets), Tokenizer (extract → train → evaluate the Chronos BPE), Train (stages in order, each with its own data; optional lazy-expert probe at every validation), Monitor (read-only view of any run folder, including runs started from the command line), Inference (lazy experts vs full memory on one prompt), Export, Serve (start and stop chronos serve), Benchmark, Auto-tune, IO metrics.

• Languages. zh-Hans, zh-Hant, en and ja; the page follows the browser language and falls back to English. Every setting has hover help with its meaning and recommended value (after a 2 s rest; the "?" mark after 0.4 s).
• Paths. "Browse…" lists folders on the server's machine. "Upload…" uses the system file manager: on the server's own machine the server opens the native dialog and the file is used where it is; from another machine the files are uploaded to <workdir>/uploads/ and that path is used. Dataset fields take several files or folders, one per line; training and export calibration merge them into one JSONL file.
• LAN access (--lan). Other machines sign in with CHRONOS_UI_KEY, or with a key printed at start when none is set; the server's own machine needs no sign-in. A signed-in browser can train and read or write files with the server user's rights, so use it on a trusted network only. The key crosses plain HTTP in clear text unless --ssl-certfile / --ssl-keyfile or a TLS proxy is used. --max-upload-mb caps a single upload (default 20480); --no-native-dialog turns the native dialog off, for example over SSH. --allow-remote with a key in the environment binds any other host.

Presets and the auto planner

The Config tab ships five v3 presets. Each one is generated by the planner and stored in ui/presets_v3.json; regenerate the file with python -m ui.presets. Parameter counts are measured with a meta-device model; peak memory is an estimate.

Preset	Total / active params (measured)	Shape	Memory level	Est. training peak
Chronos-v3 Tiny (smoke / CI)	1.51M / 1.29M	E4 top-2, vocab 6400, seq 128, depth 2	none	2.13 GB
Chronos-v3 50M (16k vocab)	49.69M / 12.09M	h384, L6, E36 top-2, vocab 16384, seq 512	none	16.4 GB (the live run measures 8.72 GB peak)
Chronos-v3 200M (32 GB) — default	200.72M / 28.20M	h576, E32 top-2	balanced	10.14 GB
Chronos-v3 500M (32 GB)	499.53M / 61.59M	E48 top-3	aggressive	9.26 GB
Chronos-v3 1B (64 GB)	1002.17M / 99.74M	E64 top-4, seq 2048	aggressive	18.92 GB

The old preset names (Recommended-CN, Tiny/Small/Medium/Large) still load; they map to the closest v3 tier.

Auto planner (Config tab → "Auto plan"). Inputs:

• target total parameters, and whether to use MoE;
• the expert count, plus either the number of active experts or an active-parameter target;
• a memory-saving level: none, balanced, aggressive or extreme;
• optional dataset paths, RAM, token budget, vocabulary size and sequence length.

chronos.model.planner.plan_training then does the following:

1. Profiles the datasets. It reads the shard manifest or samples the files, classifies records as pretrain, SFT, preference or RL, and reads the tokenizer report. The vocabulary size, sequence length (p95 / cap), token count and tokenizer path come from the data when available.
2. Searches the architecture. It looks for the hidden size, layer count, expert width and GQA heads whose meta-device count is closest to the target.
3. Sets LR and batch size from the DeepSeek LLM compute-optimal fits (η = 0.3118·C^−0.125, B = 0.2920·C^0.3271), capped for wide models.
4. Fits memory. It picks micro-batch, accumulation, int8 KV, router scope, activation checkpointing, optimizer paging and an expert RAM budget for the memory level.
5. Writes per-stage values. Pretrain, SFT, DPO, ORPO, GRPO and distill each get their own learning rate, batch, epochs, save interval and stage-specific terms (β, temperature, α).
6. Sets the loop knobs (depth above 1). recurrent_order="block", recurrent_dt_param="sigmoid" and recurrent_dt_init = min(0.5, 2 / depth), each with a sources entry in the report (the first two cite the 15M measurements; the step size is labelled estimated). One loop iteration is an Euler step of dh/dτ = block(h) − h, so depth × dt is the integration horizon; the planner holds it at 2 and lets depth refine the integration. Kernel defaults are unchanged (layer, clamp, 0.25). All of this was measured on one seed per arm at 15M parameters; see the CHANGELOG, "Physical mechanism" and "Known limits".

The stage table is editable; the Train tab merges the selected stage's row over the shared values, and the pretrain row stays in sync with the shared widgets. "Save plan" writes the config, stage values and report to JSON. Every number in the report is labelled measured or estimated.

Built-in i18n: zh-Hans, zh-Hant, en, ja

Training monitor (console tab + standalone page/API)

A chronos longrun run directory holds run_spec.json, metrics.jsonl, status.json (heartbeat), runtime.json (precision policy, device, versions, code hash), parameters.jsonl (per-tensor weight/grad/update statistics and per-virtual-layer routing), samples.jsonl, inspections/ and checkpoints. The console's Training Monitor tab and a standalone page both read these files and never touch the training device. Runs started before telemetry existed still show progress; their status is derived from metrics.jsonl and the process table.

chronos monitor --runs-dir runs                    # http://127.0.0.1:7870/
chronos serve --models-dir exports --monitor-runs-dir runs   # same page at /monitor/

JSON API (read-only): GET /api/state?run=&history=1, /api/metrics?run=&phase=&since=, /api/parameters, /api/tensors, /api/tensor?name=, /api/log?bytes=, /api/runs.

Security: the default bind is 127.0.0.1 with no authentication. Any other address requires --allow-remote and --api-key (or CHRONOS_MONITOR_KEY). Browsers sign in once and get an HttpOnly, SameSite=Strict session cookie; scripts send Authorization: Bearer <key>. Failed logins are rate-limited. Off loopback, add --ssl-certfile/--ssl-keyfile or a TLS reverse proxy, otherwise the key crosses the network in clear text.

Ollama / OpenAI protocol server

chronos serve runs the native Chronos runtime behind the Ollama and OpenAI HTTP APIs. The kernel is not converted to another architecture, so recurrent depth, SSD-lazy experts with lookahead prefetch, compressed KV and self-speculative decoding stay active.

pip install "Project_Chronos[serve]"
chronos serve --models-dir ./exports --port 11435        # 11434 would collide with a running Ollama
OLLAMA_HOST=127.0.0.1:11435 ollama list
OLLAMA_HOST=127.0.0.1:11435 ollama run <model-name> "Hello"

• Endpoints: /api/version, /api/tags, /api/show, /api/ps, /api/pull, /api/generate and /api/chat (streaming NDJSON), /v1/models, /v1/chat/completions (SSE).
• Models are <name>.pth + <name>.config.json pairs found under --models-dir.
• Chronos options go in the request options: chronos_speculative, chronos_draft_depth, chronos_gamma, chronos_miss_policy.
• Stock ollama / llama.cpp cannot execute Chronos GGUF files (MLA+SWA hybrid layers and the contractive recurrence have no stock equivalent); use this server instead.
• Security: binds 127.0.0.1 without authentication by default. Other addresses require --allow-remote and --api-key; clients then send Authorization: Bearer <key>.

Deployment export

chronos export \
    --model_path ./out/sft_384_moe.pth \
    --output_dir ./exports/sft_384 \
    --formats fp16-safetensors q8_0-safetensors fp16-gguf q8_0-gguf

Exports include config.json, chronos_export_manifest.json, and Chronos metadata for MoE top-k, shared fallback experts, lookahead router, hybrid attention, and optional expert-cache layout. Chronos can load exported safetensors/GGUF artifacts through its native lazy expert loader.

Compatibility note: the GGUF files use general.architecture=chronos. Stock Ollama/llama.cpp builds need a Chronos architecture adapter to execute them correctly; Chronos is not a LLaMA tensor-layout clone.

Stage 1: pretrain

python train_chronos.py \
    --data_path ./tests/fixtures/tiny_pretrain.jsonl \
    --hidden_size 256 --num_hidden_layers 4 --num_experts 4 \
    --epochs 1 --device cpu --save_dir ./out

Stage 2-5: alignment chain

python train_chronos_sft.py  --data_path ./tests/fixtures/tiny_sft.jsonl  --from_weight chronos --save_dir ./out --device cpu
python train_chronos_dpo.py  --data_path ./tests/fixtures/tiny_dpo.jsonl  --from_weight sft     --save_dir ./out --device cpu
python train_chronos_orpo.py --data_path ./tests/fixtures/tiny_dpo.jsonl  --from_weight sft     --save_dir ./out --device cpu
python train_chronos_grpo.py --data_path ./tests/fixtures/tiny_grpo.jsonl --from_weight orpo    --save_dir ./out --device cpu \
    --reward toy   # or lm:/path/to/reward-model

Stage 6: distillation

python train_chronos_distill.py \
    --data_path ./tests/fixtures/tiny_sft.jsonl \
    --teacher_path ./out/sft_192_moe.pth \
    --from_weight grpo --save_dir ./out --device cpu \
    --alpha 0.7 --temperature 4.0

Checkpoint and offload diagnostics

Every new .pth checkpoint writes a sibling *.config.json with the MoE topology that cannot be recovered from tensor shapes, including num_experts_per_tok. Use the diagnostic command to verify chat-template generation, no-mask vs all-available masked drift, cold shared fallback, LookaheadRouter prediction quality, and SSD/RAM/VRAM offload stats.

python diagnose_checkpoint.py \
    --model_path ./out/sft_384_moe.pth \
    --config_path ./chronos_config.json \
    --sft_data ./Dataset/sft_t2t.jsonl \
    --mlx_parity \
    --device cpu

# or through the unified CLI:
chronos diagnose --model_path ./out/sft_384_moe.pth --config_path ./chronos_config.json

For backend speed and dtype sanity checks:

python benchmark_training_backends.py --backends cpu mps mlx --dtypes auto bfloat16 float16 --steps 2

End-to-end comparison (minimind vs Chronos)

minimind is an external baseline here; pass your own checkout (Chronos does not download it).

python tools/compare_minimind_chronos_v3.py --minimind-dir /path/to/minimind \
    --pretrain_steps 150 --align_steps 30 --distill_steps 30 \
    --simulated_ssd_ms 30 --device cpu \
    --output results/compare_results_v3.json

This emits per-stage loss, HF roundtrip logit delta, tokens/sec, active-expert ratio, resident-expert bytes, M3 pipeline slack, and the backend inventory on the current host.

Cluster-pack expert storage for sequential SSD reads

from chronos.io.cluster_layout import (
    collect_activation_log,
    build_cooccurrence_matrix,
    try_louvain_clustering,
    repack_expert_weights_safetensors,
)

log = collect_activation_log(model, calib_loader, "cpu", max_batches=50)
clusters = try_louvain_clustering(build_cooccurrence_matrix(log, num_experts))
repack_expert_weights_safetensors(model, clusters, "./expert_cache_clustered")

Search lambdas and structure hyperparameters automatically

from chronos.tuning.chronos_auto_tuner import ChronosAutoTuner, ChronosSearchSpaceConfig

tuner = ChronosAutoTuner()
tuner.start(
    model_id="./out/chronos_256_moe.pth",
    dataset_path="./dataset/train.jsonl",
    search_space=ChronosSearchSpaceConfig(
        tune_lambda_balance=True,
        tune_lambda_temporal=True,
        tune_lambda_lookahead=True,
        tune_lookahead_steps=True,
        tune_hidden_size=True,
        tune_num_experts=True,
        tune_num_shared_experts=True,
        tune_kv_latent_dim=True,
        tune_recurrent_order=True,      # probes a looped model (probe_recurrent_depth, default 4)
        tune_recurrent_dt_init=True,
    ),
    n_trials=20,
)

The probe trains on real tokens from dataset_path (up to 400k, encoded with the model's tokenizer, or the bundled 32k one) and falls back to random ids with a log line when the data cannot be read. The two loop knobs are a coarse screen: a few steps of a 4-layer model cannot rank step sizes that differ by less than the run-to-run noise. recurrent_later_groups is not searched, because it trades quality for cost and a loss-only search would always choose the full block.

---

Project layout

Project_Chronos/
├── chronos/
│   ├── deps.py                    # Shim: default tokenizer path (no downloads)
│   ├── __init__.py                # AutoConfig / AutoModelForCausalLM registration
│   ├── model/
│   │   ├── kernel.py              # Native base: ChronosBaseConfig, shared RoPE/YaRN tables, FFN, GlobalAttention, logging, resume state
│   │   ├── config.py              # ChronosConfig (v3 fields, v2 migration)
│   │   ├── recurrent.py           # Loop/group embeddings, contractive step, tensor-only ACT
│   │   ├── precision_ops.py       # RMSNorm/SDPA/RoPE with one compute dtype
│   │   ├── factory.py             # Parameter-budget config search
│   │   ├── hybrid_attention.py    # MLAAttention + SlidingWindowAttention
│   │   ├── lookahead_router.py    # Per-token lookahead predictor
│   │   ├── moe_chronos.py         # ChronosMOEFeedForward + shared experts + soft gating
│   │   ├── model_chronos.py       # ChronosForCausalLM
│   │   ├── temporal_loss.py       # Temporal locality + lookahead losses
│   │   └── hf_io.py               # save/load_chronos_pretrained + HF registration
│   ├── io/
│   │   ├── expert_store.py        # Three-tier storage + per-expert events
│   │   ├── async_prefetcher.py    # Async prefetch engine
│   │   ├── storage.py             # ClusterStorage: .ctsr safetensors + manifest
│   │   ├── cluster_layout.py      # Co-occurrence clustering + repacking
│   │   └── io_simulator.py        # CHRONOS_SIM_SSD_MS test hook
│   ├── router/
│   │   ├── intent_classifier.py   # Prompt-level expert predictor
│   │   ├── expert_predictor.py    # IntentVector -> ExpertSet
│   │   └── prefill_scheduler.py   # Prefill-time expert preloader
│   ├── mlx/
│   │   ├── attention.py / moe.py / model.py / expert_store.py / inference.py
│   │   ├── moe_kernels.py         # Sorted gather_mm dispatch (mlx >= 0.32.4)
│   │   ├── speculative.py         # MLX self-speculative decoding (Torch parity)
│   │   ├── checkpointing.py       # Activation recomputation with replayed draws
│   │   └── training/              # longrun.py, compiled.py (mx.compile step), telemetry.py, recipes/
│   ├── runtime/
│   │   ├── speculative.py         # Recurrent-depth self-speculative decoding
│   │   ├── cache_manager.py       # prefetch_for_next_step / ensure_resident
│   │   ├── inference_engine.py    # End-to-end inference engine
│   │   └── metrics.py             # MetricsBus for the IO Monitor
│   ├── trainer/
│   │   ├── loss_mixin.py          # chronos_loss_term + router_kl_anchor
│   │   ├── chronos_trainer.py     # Pretrain
│   │   ├── sft_trainer.py         # Stage 2
│   │   ├── dpo_trainer.py         # Stage 3
│   │   ├── orpo_trainer.py        # Stage 4
│   │   ├── grpo_trainer.py        # Stage 5
│   │   ├── distill_trainer.py     # Stage 6
│   │   └── reward.py              # ToyReward / LMRewardModel / build_reward_fn
│   ├── tuning/
│   │   └── chronos_auto_tuner.py  # Optuna lambda + architecture search
│   ├── eval/
│   │   ├── io_profiler.py         # Lookahead validation
│   │   └── benchmark.py           # End-to-end benchmarking
│   ├── data/
│   │   ├── flexible_dataset.py    # Flexible JSONL dataset loader
│   │   └── shards.py              # memory-mapped pretrain / SFT shard readers
│   ├── tokenizer/                 # Chronos BPE builder (extract, normalize, train, evaluate)
│   ├── monitor/                   # Read-only run monitor: snapshot, API server, static page
│   ├── webui/                     # Web UI: FastAPI server, field schema, i18n/*.json, static/ SPA
│   ├── workbench/                 # Web UI jobs: train worker, data scan, file picking / merge, lazy probe
│   ├── backend/
│   │   ├── __init__.py            # BackendDispatcher (cpu/mps/cuda/mlx)
│   │   ├── dispatcher.py          # Capability probing + priority logic
│   │   └── ext/opencl.py          # Third-party OpenCL extension hook
│   ├── _backend_legacy.py         # Backward-compatible older APIs
│   ├── serving/
│   │   ├── ollama_server.py       # chronos serve: Ollama/OpenAI protocol layer
│   │   ├── runtime.py             # Model discovery, lazy experts, speculative generation
│   │   └── vllm_adapter.py        # Optional vLLM registration
│   └── cli.py                     # Unified CLI
├── ui/                            # Legacy Gradio console (chronos-ui --legacy)
│   ├── i18n.py
│   ├── estimator.py               # Live parameter / memory estimator
│   └── tabs/
│       ├── config_tab.py          # Config + Designer merged together
│       ├── data_tab.py
│       ├── train_tab.py           # Owns data_path
│       ├── monitor_tab.py         # Training Monitor (shares chronos.monitor)
│       ├── pipeline_tab.py        # Per-stage datasets across all 6 stages
│       ├── inference_tab.py
│       ├── benchmark_tab.py       # Markdown table + gr.BarPlot
│       ├── autotune_tab.py        # Persistent logs + Apply Best -> Config
│       └── iomon_tab.py           # MetricsBus dashboard
├── chronos_app.py                 # Web UI entry
├── train_chronos.py               # Stage 1 entry
├── train_chronos_sft.py           # Stage 2 entry
├── train_chronos_dpo.py           # Stage 3 entry
├── train_chronos_orpo.py          # Stage 4 entry
├── train_chronos_grpo.py          # Stage 5 entry
├── train_chronos_distill.py       # Stage 6 entry
├── tools/
│   ├── compare_minimind_chronos.py
│   ├── compare_minimind_chronos_v2.py
│   └── compare_minimind_chronos_v3.py
├── tests/
│   ├── test_smoke.py / test_smoke_cuda.py
│   ├── test_v3_kernel.py / test_speculative.py / test_moe_dispatch.py / test_precision_chain.py
│   ├── test_tokenizer_builder.py / test_longrun.py / test_monitor.py
│   └── fixtures/
├── docs/
│   └── vllm_integration.md
├── demos/
│   └── translation/               # demo extension (chronos-translate): corpus, runs, bundles, the 50M demo model; not in the wheel
├── pyproject.toml
└── README.md / README_zh.md / THIRD_PARTY_NOTICES.md

---

Roadmap

timeline
    title Project Chronos delivery map
    Phase 1
        : LookaheadRouter
        : Temporal locality regularization
        : Router-probability collection path
    Phase 2
        : Async IO engine
        : Three-tier SSD/RAM/VRAM storage
        : Co-activation clustering
    Phase 3
        : Hybrid MLA + SlidingWindow attention
        : PrefillScheduler
        : Dual-layer routing
    Phase 4
        : Native MLX backend
        : Web UI and CLI
        : Optuna search
        : Open-source release
    M1-M3
        : Cluster-aware safetensors storage
        : Supervised lookahead loss
        : Dual-stream decode with per-expert events
    M4-M6
        : SFT / DPO / ORPO / GRPO trainers
        : Router KL anchor
        : HF IO, vLLM adapter, multi-backend dispatch
        : Stage 6 distillation and pluggable rewards
        : Web UI v2, benchmark plots, IO Monitor
    v3 (0.3.0)
        : Recurrent-depth kernel with shared_first KV
        : top-2 + shared expert, grouped MoE dispatch
        : Self-speculative decoding (Torch and MLX)
        : Ollama/OpenAI protocol server
        : Ready-to-run model bundles (chronos bundle)
        : Chronos tokenizer, corpus pipeline, long-run trainer
        : Training monitor page, API and console tab
        : int8 KV, MLA absorbed decode, shared router scope (opt-in)
        : Auto planner, v3 presets, tokenizer and serve tabs
    Next
        : Multi-seed and larger-size runs of the 0.3.3 loop settings
        : Speculative decoding: a draft path that pays off on larger models and lazy experts
        : Train IntentClassifier on large activation corpora
        : Benchmark 7B+ checkpoints
        : Inject masks on the vLLM worker path
        : Ship real Vulkan / OpenCL kernels

Loading

mindmap
  root((Chronos innovation surface))
    Predictive routing
      IntentClassifier
        Prompt-level hot expert prior
        Budgeted expert-set prediction
      LookaheadRouter
        Per-token future routing
        Soft-target CE from real future routers
    Storage-aware MoE
      Clustered safetensors
        .ctsr packed expert clusters
        manifest-driven mmap
      Three-tier cache
        NVMe SSD
        Pinned RAM
        VRAM / unified memory
      Soft fallback
        Shared experts always resident
        No hard stall on cache miss
    Decode pipeline
      Prefill-time expert loading
      AsyncPrefetcher queue
      H2D stream
      Per-expert CUDA events
    Training stack
      Pretrain
      SFT
      DPO
      ORPO
      GRPO
      Distill
      Router KL anchor
    Deployment
      HF safetensors IO
      AutoModel registration
      vLLM adapter
      CPU / CUDA / MPS / MLX dispatch
      Web UI and CLI

Loading

---

Citation

@misc{chronos2026,
  title  = {Project Chronos: Prefill-Time Expert Loading and Dual-Layer Routing
             for Zero-Stall On-Device MoE Inference},
  author = {Fona and Project Chronos Contributors},
  year   = {2026},
  url    = {https://github.com/FonaTech/Project_Chronos}
}

---

Third-party attribution

Project Chronos started from jingyaogong's minimind (Apache-2.0). Since 0.3.2 no minimind code is used or downloaded. The native kernel keeps minimind's config field names and RoPE layout for checkpoint compatibility, and the legacy 6400-token tokenizer ships as data files with their licence. Full attribution lives in THIRD_PARTY_NOTICES.md.

---

License

Apache 2.0 - see LICENSE