curnane-lab · curnane-lab · Aug 27, 2026 · Jul 30, 2026 · Jul 30, 2026 · Aug 1, 2026
diff --git a/.github/workflows/publish_pypi.yaml b/.github/workflows/publish_pypi.yaml
@@ -3,31 +3,56 @@ name: Publish to PyPI
 on:
   workflow_dispatch:
 
+permissions:
+  contents: read
+
+concurrency:
+  group: publish-pypi
+  cancel-in-progress: false
+
 jobs:
-  build-n-publish:
-    if: github.event_name == 'workflow_dispatch'
-    name: Build and publish Python distributions to PyPI
+  build:
+    name: Build Python distributions
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    steps:
+      - uses: actions/checkout@v7
+        with:
+          persist-credentials: false
+
+      - uses: actions/setup-python@v7
+        with:
+          python-version: "3.11"
+
+      - name: Build and validate distributions
+        run: |
+          python -m pip install --upgrade build twine
+          python -m build
+          python -m twine check --strict dist/*
+
+      - name: Upload distributions
+        uses: actions/upload-artifact@v7
+        with:
+          name: python-distributions
+          path: dist/
+          if-no-files-found: error
+
+  publish:
+    name: Publish distributions to PyPI
+    needs: build
     runs-on: ubuntu-latest
     timeout-minutes: 20
     environment:
       name: pypi
-      url: https://pypi.org/p/specforgeee
+      url: https://pypi.org/p/specforge
     permissions:
       id-token: write
     steps:
-    - uses: actions/checkout@v2
-
-    - uses: actions/setup-python@v2
-      with:
-        python-version: '3.11'
-
-    - run: pip install build && python -m build --sdist
-
-    # publish to PyPI if executed on the main branch
-    - name: Publish package to PyPI
-      id: publish
-      uses: pypa/gh-action-pypi-publish@release/v1
-      with:
-        user: __token__
-        password: ${{ secrets.PYPI_TOKEN }}
-        verbose: true
+      - name: Download distributions
+        uses: actions/download-artifact@v8
+        with:
+          name: python-distributions
+          path: dist/
+
+      - name: Publish distributions to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml
@@ -13,7 +13,24 @@ permissions:
   contents: read
 
 jobs:
+  cleanup-runner:
+    if: (github.repository == 'sgl-project/SpecForge' || github.event_name == 'pull_request') &&
+        github.event.pull_request.draft == false
+    runs-on: [self-hosted]
+    concurrency:
+      group: specforge-gpu-ci
+      cancel-in-progress: false
+      queue: max
+    timeout-minutes: 5
+    steps:
+      # Job containers need a fresh Docker network before any container step can
+      # run, so stale networks must be pruned in a host-side prerequisite job.
+      - name: Prune stale Docker networks
+        shell: bash
+        run: docker network prune --force --filter "until=1h"
+
   unit-test:
+    needs: cleanup-runner
     if: (github.repository == 'sgl-project/SpecForge' || github.event_name == 'pull_request') &&
         github.event.pull_request.draft == false
     runs-on: [self-hosted]

diff --git a/README.md b/README.md
@@ -16,11 +16,12 @@
 SpecForge is an ecosystem project developed by the SGLang team. It is a framework for training speculative decoding models so that you can smoothly port them over to the SGLang serving framework to speed up your inference.
 
 We have seen many open-source projects for speculative decoding, but most of them are not well-maintained or not directly compatible with SGLang. We prepared this project because we wish that the open-source community can enjoy a speculative decoding framework that is
+
 - regularly maintained by the SpecForge team: the code is runnable out-of-the-box
-- directly compatible with SGLang: there is no additional efforts for porting to SGLang
-- provides local offline and server-only online-disaggregated training through
-  one runtime, including the supported data, tensor, and sequence parallel
-  topologies
+- directly compatible with SGLang: no additional porting effort is required
+- able to run online disaggregated training and both colocated and
+  disaggregated offline training through one runtime, including the supported
+  data, tensor, and sequence parallel topologies
 
 
 Check out [**our documentation**](https://docs.sglang.ai/SpecForge/) to get started.
@@ -31,23 +32,25 @@ Check out [**our documentation**](https://docs.sglang.ai/SpecForge/) to get star
 Every method uses the same typed training entry point:
 
 ```bash
-specforge train --config examples/configs/qwen3-8b-eagle3-disaggregated.yaml
+specforge train --config examples/configs/online/disaggregated/external/qwen3-8b-eagle3-disaggregated.yaml
 ```
 
-The typed `deployment.trainer` topology self-launches trainer DP and EAGLE3
-offline USP process groups. A single-node disaggregated config also supervises
-its SpecForge producer and consumer; Mooncake and SGLang remain externally
-managed services, and online target parallelism belongs to SGLang. There are no
-method-specific Python training entry points.
+The path under `examples/configs` identifies feature mode, topology, and online
+service ownership. The command above uses an `external` recipe: SpecForge
+supervises the producer and consumer on one trainer node, while the user or
+scheduler owns Mooncake and SGLang. Recipes under `managed-local` also start
+those services on the local host. Online target parallelism belongs to SGLang;
+`deployment.trainer` owns trainer DP and offline EAGLE3 USP process groups.
+There are no method-specific Python training entry points.
 
 | Method | Description | Example config | Optimization |
 | --- | --- | --- | --- |
-| **[EAGLE3](https://arxiv.org/abs/2503.01840)** | Feature-based autoregressive drafting | [Online](./examples/configs/qwen3-8b-eagle3-disaggregated.yaml) / [Offline](./examples/configs/qwen3-8b-eagle3-offline.yaml) / [Disaggregated offline](./examples/configs/qwen3-8b-eagle3-offline-disaggregated.yaml) | [LK loss](https://arxiv.org/pdf/2602.23881) |
-| **[P-EAGLE](https://arxiv.org/abs/2602.01469)** | Parallel EAGLE | [Online](./examples/configs/qwen3-8b-peagle-disaggregated.yaml) | — |
-| **EAGLE3.1** | Feature-based autoregressive drafting with attention drift | [Online](./examples/configs/qwen3-30b-a3b-eagle3.1-online.yaml) | — |
-| **[DFlash](https://arxiv.org/abs/2602.06036)** | Block-parallel drafting | [Online](./examples/configs/qwen3-8b-dflash-online.yaml) / [Disaggregated](./examples/configs/qwen3-8b-dflash-disaggregated.yaml) | [D-PACE](https://arxiv.org/abs/2605.18810) |
-| **[Domino](https://arxiv.org/html/2605.29707v1)** | DFlash with GRU logit correction | [Online](./examples/configs/qwen3-8b-domino-online.yaml) / [Disaggregated](./examples/configs/qwen3-8b-domino-disaggregated.yaml) | — |
-| **[DSpark](https://arxiv.org/abs/2607.05147)** | Confidence-Scheduled Semi-Autoregressive Generation | [Disaggregated](./examples/configs/qwen3-4b-dspark-disaggregated.yaml) | — |
+| **[EAGLE3](https://arxiv.org/abs/2503.01840)** | Feature-based autoregressive drafting | [Online external](./examples/configs/online/disaggregated/external/qwen3-8b-eagle3-disaggregated.yaml) / [Offline colocated](./examples/configs/offline/colocated/qwen3-8b-eagle3-offline.yaml) / [Offline disaggregated](./examples/configs/offline/disaggregated/qwen3-8b-eagle3-offline-disaggregated.yaml) | [LK loss](https://arxiv.org/pdf/2602.23881) |
+| **[P-EAGLE](https://arxiv.org/abs/2602.01469)** | Parallel EAGLE | [Online external](./examples/configs/online/disaggregated/external/qwen3-8b-peagle-disaggregated.yaml) | — |
+| **EAGLE3.1** | Feature-based autoregressive drafting with attention drift | [Online external](./examples/configs/online/disaggregated/external/qwen3-30b-a3b-eagle3.1-online.yaml) | — |
+| **[DFlash](https://arxiv.org/abs/2602.06036)** | Block-parallel drafting | [Online external](./examples/configs/online/disaggregated/external/qwen3-8b-dflash-online.yaml) / [Offline colocated](./examples/configs/offline/colocated/qwen3-8b-dflash-offline.yaml) / [Online managed-local](./examples/configs/online/disaggregated/managed-local/qwen3-8b-dflash-1server-dp7-disaggregated.yaml) | [D-PACE](https://arxiv.org/abs/2605.18810) |
+| **[Domino](https://arxiv.org/html/2605.29707v1)** | DFlash with GRU logit correction | [Online external](./examples/configs/online/disaggregated/external/qwen3-8b-domino-online.yaml) / [Offline colocated](./examples/configs/offline/colocated/qwen3-8b-domino-offline.yaml) / [Online managed-local](./examples/configs/online/disaggregated/managed-local/qwen3-8b-domino-multiserver-disaggregated.yaml) | — |
+| **[DSpark](https://arxiv.org/abs/2607.05147)** | Confidence-Scheduled Semi-Autoregressive Generation | [Online external](./examples/configs/online/disaggregated/external/qwen3-4b-dspark-disaggregated.yaml) / [Offline colocated](./examples/configs/offline/colocated/qwen3-4b-dspark-offline.yaml) | — |
 
 See the [training guide](./docs/basic_usage/training.md) for the supported
 method/topology matrix and the
@@ -69,11 +72,14 @@ SpecBundle is a collection of production-grade speculative decoding models that
 
 
 ## 🎉 News
-
+- [2026-08] 🎉 Released SpecBundle (phase 2) and SpecForge v0.3.0. Check out our blog at [LMSYS.org](https://www.lmsys.org/blog/2026-08-04-specforge-v0-3)
+- [2026-07] 🚀 Day0 supported two flagship dspark draft model, [Inkling](https://huggingface.co/RadixArk/Inkling-DSpark-Preview) and [Kimi-K3](https://huggingface.co/RadixArk/Kimi-K3-DSpark).
+- [2026-07] 🔥 Supported full disaggregation of training and inference in online training.
+- [2026-07] 🔥 Added DSpark online training for DFlash draft models.
 - [2026-06] 🔥 Added D-PACE as an optional loss for DFlash training.
 - [2026-06] 🔥 Added Domino online training for DFlash draft models.
 - [2026-01] 🔥 Added DFlash block-parallel online training with SGLang serving support.
-- [2025-12] 🎉 Released SpecBundle (phase 1) and SpecForge v0.2. Check out our blog at [LMSYS.org](https://lmsys.org/blog/2025-12-23-spec-bundle-phase-1/)
+- [2025-12] 🎉 Released SpecBundle (phase 1) and SpecForge v0.2.0. Check out our blog at [LMSYS.org](https://lmsys.org/blog/2025-12-23-spec-bundle-phase-1/)
 - [2025-08] 🔔 SpecForge is listed as a [flagship project](https://lmsys.org/about/) in LMSYS. Congratulations to the SpecForge team!
 - [2025-08] 🔥 SpecForge powered the Eagle3 draft model for GPT-OSS. Check out the blog at [LMSYS.org](https://lmsys.org/blog/2025-08-27-gpt-oss/)
 - [2025-07] 🔥 SpecForge is released together with Llama4-Eagle3 checkpoints. Check out our blog at [LMSYS.org](https://lmsys.org/blog/2025-07-25-spec-forge/)

diff --git a/configs/kimi-k3-dspark.json b/configs/kimi-k3-dspark.json
@@ -0,0 +1,53 @@
+{
+  "architectures": ["DSparkDraftModel"],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {"AutoModel": "dspark.DSparkDraftModel"},
+  "block_size": 7,
+  "bos_token_id": 163584,
+  "dflash_config": {
+    "attention_mode": "gqa",
+    "confidence_head_alpha": 1.0,
+    "confidence_head_with_markov": true,
+    "enable_confidence_head": true,
+    "markov_head_type": "vanilla",
+    "markov_rank": 256,
+    "mask_token_id": 163824,
+    "projector_type": "dspark",
+    "target_layer_ids": [7, 23, 51, 67, 83]
+  },
+  "dtype": "bfloat16",
+  "eos_token_id": 163586,
+  "head_dim": 64,
+  "hidden_act": "silu",
+  "hidden_size": 7168,
+  "initializer_range": 0.02,
+  "intermediate_size": 14336,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 65536,
+  "max_window_layers": 5,
+  "model_type": "qwen3",
+  "num_attention_heads": 64,
+  "num_hidden_layers": 5,
+  "num_key_value_heads": 16,
+  "num_target_layers": 93,
+  "pad_token_id": 163839,
+  "rms_norm_eps": 1e-05,
+  "rope_parameters": {
+    "factor": 16.0,
+    "original_max_position_embeddings": 65536,
+    "rope_theta": 10000.0,
+    "rope_type": "yarn"
+  },
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 163840
+}
diff --git a/configs/qwen3.5-4b-vl-dflash.json → configs/qwen3.5-4b-mtp.json b/configs/qwen3.5-4b-vl-dflash.json → configs/qwen3.5-4b-mtp.json
@@ -1,50 +1,37 @@
 {
   "architectures": [
-    "DFlashDraftModel"
+    "Qwen3_5MTPDraftModel"
   ],
   "attention_bias": false,
   "attention_dropout": 0.0,
   "auto_map": {
-    "AutoModel": "dflash.DFlashDraftModel"
+    "AutoModel": "mtp.Qwen3_5MTPDraftModel"
   },
-  "block_size": 16,
+  "attn_output_gate": true,
   "bos_token_id": 248043,
-  "dflash_config": {
-    "mask_token_id": 248070,
-    "target_layer_ids": [1, 8, 15, 22, 29]
-  },
-  "dtype": "bfloat16",
   "eos_token_id": 248044,
   "head_dim": 256,
   "hidden_act": "silu",
   "hidden_size": 2560,
   "initializer_range": 0.02,
   "intermediate_size": 9216,
-  "layer_types": [
-    "full_attention",
-    "full_attention",
-    "full_attention",
-    "full_attention",
-    "full_attention"
-  ],
   "max_position_embeddings": 262144,
-  "model_type": "qwen3_vl_text",
+  "model_type": "qwen3",
+  "mtp_config": {
+    "share_lm_head": true
+  },
   "num_attention_heads": 16,
-  "num_hidden_layers": 5,
+  "num_hidden_layers": 1,
   "num_key_value_heads": 4,
-  "num_target_layers": 32,
   "pad_token_id": 248044,
   "partial_rotary_factor": 0.25,
   "rms_norm_eps": 1e-06,
-  "rope_scaling": {
-    "mrope_interleaved": true,
-    "mrope_section": [11, 11, 10],
-    "rope_type": "default",
-    "partial_rotary_factor": 0.25
-  },
+  "rope_scaling": null,
   "rope_theta": 10000000,
   "sliding_window": null,
   "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
   "use_cache": true,
+  "use_sliding_window": false,
   "vocab_size": 248320
 }
diff --git a/docs/advanced_features/customization.md b/docs/advanced_features/customization.md
@@ -13,7 +13,7 @@ For one-off changes, use dotted overrides rather than adding another launcher:
 
 ```bash
 specforge train \
-  --config examples/configs/qwen3-8b-eagle3-disaggregated.yaml \
+  --config examples/configs/online/disaggregated/external/qwen3-8b-eagle3-disaggregated.yaml \
   model.target_model_path=/models/my-target \
   data.train_data_path=/datasets/my-training-data.jsonl \
   training.learning_rate=5e-5