FonaTech · Nov 13, 2026 · Nov 7, 2026 · Nov 7, 2026 · Nov 7, 2026 · Nov 7, 2026
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -4,7 +4,7 @@ on:
   push:
     branches: [main, dev]
   pull_request:
-    branches: [main]
+    branches: [main, dev]
 
 jobs:
   test:
@@ -36,3 +36,8 @@ jobs:
       - name: Unit tests
         run: |
           python -m pytest tests/ -v --tb=short
+
+      - name: Demo driver tests
+        run: |
+          # The translation demo's model driver; the kernel suite above stays task-neutral.
+          python -m pytest demos/translation/tests/test_model_enzh_50m_sft.py -v --tb=short
diff --git a/.github/workflows/python-app.yml b/.github/workflows/python-app.yml
@@ -27,7 +27,7 @@ jobs:
       run: |
         python -m pip install --upgrade pip
         pip install torch==2.10.0 --index-url https://download.pytorch.org/whl/cpu
-        pip install -c .github/ci-constraints.txt flake8 pytest psutil -r requirements.txt
+        pip install -c .github/ci-constraints.txt flake8 pytest psutil fastapi uvicorn httpx -r requirements.txt
     - name: Lint with flake8
       run: |
         # stop the build if there are Python syntax errors or undefined names

diff --git a/CHANGELOG.md b/CHANGELOG.md
diff --git a/MANIFEST.in b/MANIFEST.in
@@ -1,5 +1,7 @@
 include README.md README_zh.md LICENSE THIRD_PARTY_NOTICES.md CHANGELOG.md requirements.txt
 recursive-include chronos/monitor/static *.html *.js *.css
+recursive-include chronos/webui/static *.html *.js *.css *.svg
+recursive-include chronos/webui/i18n *.json
 include ui/event_contract.js
 global-exclude __pycache__ *.py[cod] .DS_Store
 prune runs
@@ -10,5 +12,5 @@ prune results
 prune reports
 prune .archive
 prune Translation?Training?Data
+prune demos
 prune chronos/mlx/training/recipes
-prune models
diff --git a/README.md b/README.md
diff --git a/README_zh.md b/README_zh.md
diff --git a/THIRD_PARTY_NOTICES.md b/THIRD_PARTY_NOTICES.md
@@ -13,36 +13,26 @@ below or linked.
 **License**: Apache License, Version 2.0
 **Copyright**: Copyright jingyaogong and MiniMind Contributors
 
-**Files derived from MiniMind**:
+**Status since 0.3.2: no runtime dependency, no bundled code.** Releases up to
+0.3.1a3 imported a minimind checkout at runtime (`chronos.deps` cloned
+`master` into `~/.cache/chronos/minimind-master/` and put it on `sys.path`).
+0.3.2 removes that dependency. The pieces Chronos used are re-implemented in
+`chronos/model/kernel.py` (`ChronosBaseConfig`, RoPE/YaRN tables, gated FFN,
+logging and resume utilities). Nothing is downloaded.
 
-| Chronos file | Derived from MiniMind |
-|---|---|
-| `chronos/model/config.py` | `model/model_minimind.py` — `MiniMindConfig` |
-| `chronos/model/model_chronos.py` | `model/model_minimind.py` — `MiniMindModel`, `MiniMindForCausalLM`, `MiniMindBlock` |
-| `chronos/model/moe_chronos.py` | `model/model_minimind.py` — `MOEFeedForward`, `FeedForward` |
-| `chronos/model/hybrid_attention.py` | `model/model_minimind.py` — `Attention`, `apply_rotary_pos_emb`, `repeat_kv`, `RMSNorm` |
-| `chronos/model/lookahead_router.py` | Architecture inspired by MiniMind's routing design |
-| `chronos/trainer/chronos_trainer.py` | `trainer/train_pretrain.py` — pretraining loop structure |
-| `chronos/trainer/sft_trainer.py` | `trainer/train_full_sft.py` — SFT loop + chat-template loss masking |
-| `chronos/trainer/dpo_trainer.py` | `trainer/train_dpo.py` — log-σ(β·logits) DPO formulation, `_logits_to_log_probs` helper |
-| `chronos/trainer/grpo_trainer.py` | `trainer/train_grpo.py` — GRPO group-relative PG + KL formulation |
-| `chronos/trainer/distill_trainer.py` | `trainer/train_distillation.py` — temperature-scaled KL distillation |
-| `chronos/trainer/reward.py` | `trainer/trainer_utils.py` — `LMForRewardModel` ABI (`get_score`) wrapper |
-| `chronos/eval/io_profiler.py` | `eval_llm.py` — evaluation patterns |
-
-**Nature of modifications**:
+**What still derives from MiniMind**:
 
-- `ChronosConfig` extends `MiniMindConfig` with new fields: `lookahead_steps`, hybrid attention dims, λ loss coefficients (balance / temporal / lookahead / router-anchor), VRAM budget, pinned RAM fraction, `storage_format`, `cluster_manifest_path`.
-- `ChronosForCausalLM` replaces `MiniMindForCausalLM` with hybrid attention, LookaheadRouter injection, dual-layer routing hooks, an `available_expert_masks` forward kwarg, and `_tied_weights_keys` / `all_tied_weights_keys` shims for HF safetensors round-trips.
-- `ChronosMOEFeedForward` extends `MOEFeedForward` with always-resident shared experts and compile-safe soft gating (`avail[i] * expert_out + (1−avail[i]) * shared_out`).
-- `MLAAttention` and `SlidingWindowAttention` are new implementations sharing RMSNorm and RoPE utilities from MiniMind's `Attention`.
-- The full Stage 1–6 trainer family (Pretrain / SFT / DPO / ORPO / GRPO / Distill) wraps the corresponding minimind loops where they exist (Pretrain / SFT / DPO / GRPO / Distill) and adds **ORPO from scratch** (Hong et al. 2024 — minimind has no ORPO trainer). Every wrapper injects `chronos_loss_term` (lookahead + temporal + balance) and an optional `router_kl_anchor` term.
-- `ChronosGRPOTrainer` is **self-contained**: it ships its own rollout (vanilla `model.generate` style sampling) and accepts a pluggable `reward_fn`. It does not require minimind's sglang-based `rollout_engine`.
+| Chronos file | What it keeps from MiniMind |
+|---|---|
+| `chronos/model/kernel.py` | Field names and defaults of `MiniMindConfig`, the rotate-half RoPE/YaRN layout and numerics, and the `gate_proj/up_proj/down_proj` FFN naming. They are kept so that checkpoints trained before 0.3.2 load bit-identically. The code is a new implementation. |
+| `chronos/assets/tokenizers/legacy_6400/` | `tokenizer.json` and `tokenizer_config.json`, copied unchanged (data files only) from commit `24896cd2c4e9328a0911cf69ff61ace17a459b71`, with its `LICENSE` and a `NOTICE` that records the sha256 of each file. It is the default tokenizer only for old checkpoints that never recorded one. |
+| `chronos/trainer/sft_trainer.py`, `dpo_trainer.py`, `grpo_trainer.py`, `distill_trainer.py` | The loop structure and loss formulations first followed minimind's `train_full_sft.py`, `train_dpo.py`, `train_grpo.py` and `train_distillation.py` (chat-template loss masking, log-σ(β·Δ) DPO, group-relative policy gradient with KL, temperature-scaled KD). These trainers are self-contained Chronos code. ORPO was written from scratch (Hong et al. 2024). |
+| `chronos/trainer/reward.py` | The `get_score(tokenizer, messages)` reward-model ABI wrapper. |
+| `tools/compare_minimind_chronos*.py` | Optional comparison harnesses that run minimind as an external baseline. They need a checkout the user provides through `--minimind-dir` or `CHRONOS_MINIMIND_DIR`. |
 
-**Automatic download**: When `minimind-master` is not found locally, Project Chronos
-automatically clones it from `https://github.com/jingyaogong/minimind` into
-`~/.cache/chronos/minimind-master/`. This download is governed by the Apache-2.0
-license of the MiniMind project.
+Chronos-specific parts (MLA/SWA hybrid attention, recurrent-depth controller,
+LookaheadRouter, shared-expert MoE, lazy expert paging, speculative decoding,
+MLX backend, tokenizer builder) are original work.
 
 **Apache-2.0 License Text**:
 
@@ -60,7 +50,8 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ```
 
-Full license: https://github.com/jingyaogong/minimind/blob/master/LICENSE
+Full license: https://github.com/jingyaogong/minimind/blob/master/LICENSE (also shipped as
+`chronos/assets/tokenizers/legacy_6400/LICENSE`)
 
 ---
 

diff --git a/benchmark_compare.py b/benchmark_compare.py
@@ -1,5 +1,5 @@
 """
-benchmark_compare.py — Background benchmark: minimind (pure RAM) vs Chronos (SSD+DRAM)
+benchmark_compare.py — Background benchmark: dense baseline (pure RAM) vs Chronos (SSD+DRAM)
 
 Runs in background, writes results to benchmark_results.json
 Usage: python benchmark_compare.py
@@ -11,7 +11,6 @@
 if _pkg_root not in sys.path:
     sys.path.insert(0, _pkg_root)
 
-import chronos.deps  # auto-bootstrap minimind on sys.path
 import torch
 import psutil
 
@@ -68,18 +67,25 @@ def measure_tokens_per_sec(model, vocab_size, device, n_tokens=50, seq_len=32):
     elapsed = time.monotonic() - t0
     return n_tokens / max(elapsed, 1e-6)
 
-# ── Benchmark 1: minimind pure RAM ────────────────────────────────
+# ── Benchmark 1: dense baseline (pure RAM) ─────────────────────────
+# A conventional transformer built from the same native kernel: full causal
+# GQA (no MLA/SWA), one pass per layer (no recurrence), all experts resident.
+# Before 0.3.2 this slot ran minimind; tools/compare_minimind_chronos*.py
+# still compare against an external minimind checkout on request.
 
-def bench_minimind():
-    log('=== Benchmark 1: MiniMind (pure RAM) ===')
-    from model.model_minimind import MiniMindConfig, MiniMindForCausalLM
+def bench_baseline():
+    log('=== Benchmark 1: dense baseline (pure RAM) ===')
+    from chronos.model.config import ChronosConfig
+    from chronos.model.model_chronos import ChronosForCausalLM
 
-    cfg = MiniMindConfig(
+    cfg = ChronosConfig(
         hidden_size=512, num_hidden_layers=8, use_moe=True,
-        num_experts=4, num_experts_per_tok=1,
+        num_experts=4, num_experts_per_tok=1, num_shared_experts=0,
+        use_hybrid_attention=False, recurrent_enabled=False,
+        lambda_lookahead=0.0, lambda_temporal=0.0,
     )
     ram_before = ram_used_gb()
-    model = MiniMindForCausalLM(cfg)
+    model = ChronosForCausalLM(cfg)
     model.eval()
     ram_after = ram_used_gb()
     ram_model_gb = ram_after - ram_before
@@ -93,7 +99,7 @@ def bench_minimind():
     train_losses = []
     for step in range(5):
         x = torch.randint(0, cfg.vocab_size, (2, 64))
-        out = model(x, labels=x)
+        out, _lookahead = model(x, labels=x)
         loss = out.loss + out.aux_loss
         loss.backward()
         optimizer.step()
@@ -111,13 +117,14 @@ def bench_minimind():
     log(f'  TPS={tps:.1f}  RAM_train={ram_train:.3f}GB  RAM_infer={ram_infer:.3f}GB')
 
     result = {
-        'name': 'MiniMind (pure RAM)',
+        'name': 'Dense baseline (GQA, pure RAM)',
         'params_m': round(params_m, 1),
         'ram_model_gb': round(ram_model_gb, 3),
         'ram_train_gb': round(ram_train, 3),
         'ram_infer_gb': round(ram_infer, 3),
         'tokens_per_sec': round(tps, 1),
         'train_losses': [round(l, 4) for l in train_losses],
+        'kv_cache_type': 'GQA (full K/V)',
     }
     del model
     gc.collect()
@@ -241,9 +248,9 @@ def bench_chronos():
 
     results = {}
     try:
-        results['minimind'] = bench_minimind()
+        results['baseline'] = bench_baseline()
     except Exception as e:
-        log(f'minimind bench failed: {e}')
+        log(f'baseline bench failed: {e}')
         import traceback; log(traceback.format_exc())
 
     try:

diff --git a/chronos/__init__.py b/chronos/__init__.py
@@ -1,9 +1,5 @@
 # Project Chronos: On-device low-latency lookahead dual-layer MoE inference
-__version__ = "0.3.0"
-
-# Bootstrap minimind dependency before any submodule imports
-from chronos.deps import ensure_minimind as _ensure_minimind
-_ensure_minimind()
+__version__ = "0.3.3"
 
 # Register with HuggingFace AutoXXX factories so
 # AutoModelForCausalLM.from_pretrained("path/to/chronos") works out of the box.

diff --git a/chronos/app.py b/chronos/app.py
@@ -36,7 +36,6 @@
 from ui.gradio_compat import gr
 from ui.event_contract import UIEventContract, event_contract_app_kwargs
 
-import chronos.deps  # auto-bootstrap minimind
 from ui.i18n import (
     LANGUAGE_CHOICES, DEFAULT_LANGUAGE,
     t, set_current_lang, build_language_update, get_registered_components,
@@ -78,7 +77,7 @@ def _hero_html() -> str:
         "<span class='app-meta-chip'>Apache-2.0</span>"
         "<span class='app-meta-chip'>Python 3.10+</span>"
         "<span class='app-meta-chip'>PyTorch 2.1+</span>"
-        "<span class='app-meta-chip'><a href='https://github.com/jingyaogong/minimind' target='_blank'>MoE kernel: minimind</a></span>"
+        "<span class='app-meta-chip'>Kernel: Chronos native</span>"
         "</div>"
         "</div>"
     )

diff --git a/chronos/assets/__init__.py b/chronos/assets/__init__.py