messages listlengths 2 2 | category stringclasses 5
values | sources listlengths 1 3 | id stringlengths 25 25 |
|---|---|---|---|
[
{
"role": "user",
"content": "In Halo Configuration Reference / Performance & balancing flags, what does `enable_torch_profiler` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "Every standard training script inherits these from `CommonScriptArguments`; they... | configuration | [
{
"path": "agent-docs/reference/configuration-reference.md",
"section": "Configuration Reference / Performance & balancing flags",
"section_id": "agent-docs/reference/configuration-reference.md:L64",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/hal... | halo-00200e5b4b82df519173 |
[
{
"role": "user",
"content": "In Halo Configuration Reference / OfflineGRPOConfig, what does `kl_beta` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "`OfflineGRPOTrainer`: group-relative policy optimization on pre-computed rewards. **Source:** `src/configs... | configuration | [
{
"path": "agent-docs/reference/configuration-reference.md",
"section": "Configuration Reference / OfflineGRPOConfig",
"section_id": "agent-docs/reference/configuration-reference.md:L346",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e... | halo-003bd015e5ca82117d7f |
[
{
"role": "user",
"content": "In Halo Tensor Parallelism (TP) / Limitations, what does `use_peft` / LoRA (incl. EP+TP) control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the conf... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-004fcbd168010a3ba12f |
[
{
"role": "user",
"content": "For Halo Configuration Reference / Script arguments, what should I know about RLVROnlineGRPOScriptArguments?"
},
{
"role": "assistant",
"content": "RLVR Online GRPO. **Source:** `src/args/rlvr_online_grpo_args.py`\n\n| Parameter | Type | Default | Description |\n|--... | configuration | [
{
"path": "agent-docs/reference/configuration-reference.md",
"section": "Configuration Reference / Script arguments / RLVROnlineGRPOScriptArguments",
"section_id": "agent-docs/reference/configuration-reference.md:L587",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github... | halo-0053e6def62576c1a0d0 |
[
{
"role": "user",
"content": "How do I launch Halo Online SDPG?"
},
{
"role": "assistant",
"content": "```bash\n# vLLM on GPU 7 and the trainer on 0-6, the compose defaults (server setup: Online GRPO)\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6 torchrun --nproc_per_node=7 \\\n scripts/training/online... | commands-and-recipes | [
{
"path": "agent-docs/training-methods/distillation/online-sdpg.md",
"section": "Online SDPG / Launch",
"section_id": "agent-docs/training-methods/distillation/online-sdpg.md:L24",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963... | halo-008a187637667fb1e418 |
[
{
"role": "user",
"content": "In Halo Context Parallelism (CP) / Limitations, what does multimodal inputs (`pixel_values`) control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the ... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-008c92edde95ef46600a |
[
{
"role": "user",
"content": "For Halo Data Parallelism, what should I know about HSDP (Hybrid Sharded Data Parallel)?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\nsurface, the rank math, the trainer gates, a... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-0093aad39869d2bdc3af |
[
{
"role": "user",
"content": "For Halo Quickstart, what should I know about Other methods?"
},
{
"role": "assistant",
"content": "Same YAML shape, different script and a few method fields. `halo launch <method> <config>` runs any of them by name and picks the launcher (`halo launch --list`).\n\n... | commands-and-recipes | [
{
"path": "agent-docs/getting-started/quickstart.md",
"section": "Quickstart / 2. Other methods",
"section_id": "agent-docs/getting-started/quickstart.md:L47",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff... | halo-00944db1cdf35a9688f2 |
[
{
"role": "user",
"content": "In Halo Parallelism / Supported combinations, what are the details for TP only?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\nsurface, the rank math, the trainer gates, and the st... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-00a2ef8d617fc2201712 |
[
{
"role": "user",
"content": "How is the reference model handled in Halo Direct Preference Optimization (DPO)?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\nsurface, the rank math, the trainer gates, and the s... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-00acd307f378e5c894b5 |
[
{
"role": "user",
"content": "In Halo Flash Attention & Attention Backends / Choosing a backend, what are the details for Context Parallelism?"
},
{
"role": "assistant",
"content": "Every row is what the loader picks on its own; the reason for each redirect is in\nModel-specific handling.\n\n| S... | concepts-and-behavior | [
{
"path": "agent-docs/optimization/flash-attention.md",
"section": "Flash Attention & Attention Backends / Choosing a backend",
"section_id": "agent-docs/optimization/flash-attention.md:L200",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/... | halo-00c26a1da45c445ac6f0 |
[
{
"role": "user",
"content": "What does Halo support for Halo / Command A+ cookbook?"
},
{
"role": "assistant",
"content": "| FSDP | EP | CP | TP | ETP | EP+CP | EP+TP | LoRA |\n|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|\n| Yes | Yes | untested | untested | untested | untested | untested ... | commands-and-recipes | [
{
"path": "human-docs/cookbooks/halo-command-a-plus-cookbook.md",
"section": "Halo / Command A+ cookbook / Halo support",
"section_id": "human-docs/cookbooks/halo-command-a-plus-cookbook.md:L14",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/bl... | halo-00e5d7fd93597542737b |
[
{
"role": "user",
"content": "While using Halo, I see this symptom: **`The launcher declares world_size=N but ['MASTER_ADDR', 'MASTER_PORT'] is unset`** at startup. What causes it and how can I fix it?"
},
{
"role": "assistant",
"content": "| Symptom | Cause | Fix | Deep page |\n|---|---|---|---... | troubleshooting | [
{
"path": "agent-docs/reference/troubleshooting.md",
"section": "Troubleshooting / Symptom → cause → fix",
"section_id": "agent-docs/reference/troubleshooting.md:L22",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac31... | halo-00fc5cb8343b9bb0a226 |
[
{
"role": "user",
"content": "For Halo Padding-Free Flash Attention Collator, what should I know about Benchmark results?"
},
{
"role": "assistant",
"content": "Throughput is **real (non-padding) tokens/s/GPU** — `attention_mask.sum()`, not padded element count — set via `include_num_input_token... | concepts-and-behavior | [
{
"path": "agent-docs/optimization/padding-free-collator.md",
"section": "Padding-Free Flash Attention Collator / Benchmark results",
"section_id": "agent-docs/optimization/padding-free-collator.md:L61",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle... | halo-011183739ac818d55c26 |
[
{
"role": "user",
"content": "In Halo Metrics and Troubleshooting / Logged metrics, what are the details for `outcome/*`?"
},
{
"role": "assistant",
"content": "| Namespace | Answers | Keys to watch |\n|---|---|---|\n| `outcome/*` | task success | `solve_rate`, plus per-env keys such as `test_pa... | troubleshooting | [
{
"path": "agent-docs/training-methods/grpo/async-grpo/monitoring.md",
"section": "Metrics and Troubleshooting / Logged metrics",
"section_id": "agent-docs/training-methods/grpo/async-grpo/monitoring.md:L5",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whiteci... | halo-0122a87ab22895519602 |
[
{
"role": "user",
"content": "In Halo Scripts Reference / Post-training scripts, what are the details for `scripts/after_training/merge_models.py`?"
},
{
"role": "assistant",
"content": "| Script | Description |\n|--------|-------------|\n| `scripts/after_training/merge_models.py` | Merge same-a... | concepts-and-behavior | [
{
"path": "agent-docs/reference/scripts-reference.md",
"section": "Scripts Reference / Post-training scripts",
"section_id": "agent-docs/reference/scripts-reference.md:L167",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce752... | halo-01242abf70d42616b6df |
[
{
"role": "user",
"content": "For Halo Pre-training (from scratch & continued), what should I know about Raw-text data preparation?"
},
{
"role": "assistant",
"content": "Raw text is tokenized with document packing and no chat template:\n\n```bash\npython scripts/before_training/prepare_dataset.... | concepts-and-behavior | [
{
"path": "agent-docs/training-methods/pretraining.md",
"section": "Pre-training (from scratch & continued) / Raw-text data preparation",
"section_id": "agent-docs/training-methods/pretraining.md:L52",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/h... | halo-0134e2b0f1efae837716 |
[
{
"role": "user",
"content": "For Halo Rollout Servers, what should I know about SGLang?"
},
{
"role": "assistant",
"content": "`Dockerfile.sglang` builds `sglang-server:0.5.17` to align NCCL: upstream's wheel trails `uv.lock`'s\nexact pin (what the training images run), and weight sync needs bo... | concepts-and-behavior | [
{
"path": "agent-docs/infrastructure/rollout-servers.md",
"section": "Rollout Servers / SGLang",
"section_id": "agent-docs/infrastructure/rollout-servers.md:L451",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146cc... | halo-0187e17a794e5431d078 |
[
{
"role": "user",
"content": "How do I use Halo BF16 Master Weights with Stochastic Rounding?"
},
{
"role": "assistant",
"content": "AdamWBF16 is auto-enabled when `bf16: true` and `optim` is the default AdamW. No separate flag, no special launch flags:\n\n```bash\ntorchrun --nproc_per_node=8 \\... | commands-and-recipes | [
{
"path": "agent-docs/optimization/bf16-optimizer.md",
"section": "BF16 Master Weights with Stochastic Rounding / Usage",
"section_id": "agent-docs/optimization/bf16-optimizer.md:L38",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65... | halo-01e586965e0cb9c15656 |
[
{
"role": "user",
"content": "In Halo Architecture / Component map, what are the details for Checkpoints?"
},
{
"role": "assistant",
"content": "| Layer | Owns | Source | Deeper page |\n|---|---|---|---|\n| Checkpoints | The per-mode saver ladder, weight resume, per-rank optimizer shards + LR sc... | internals | [
{
"path": "agent-docs/reference/architecture.md",
"section": "Architecture / Component map",
"section_id": "agent-docs/reference/architecture.md:L16",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/a... | halo-021321b0c87c331bc992 |
[
{
"role": "user",
"content": "In Halo Tensor Parallelism (TP) / Limitations, what does native EP expert LoRA control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\nsurfac... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-021dbea01ba876c5cc7b |
[
{
"role": "user",
"content": "In Halo Tensor Parallelism (TP) / Limitations, what does `added_special_tokens` that grow the vocab control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-023e7b2874240998f658 |
[
{
"role": "user",
"content": "For Halo Qwen3 Family, what should I know about Qwen3-VL?"
},
{
"role": "assistant",
"content": "`Qwen3VLTextAttention` reuses `Qwen3MoeUlyssesAttention` for CP and is in the selective-TP accept-list (`src/distributed/tensor_parallel/module_types.py`). That accept-l... | concepts-and-behavior | [
{
"path": "agent-docs/models/qwen3.md",
"section": "Qwen3 Family / Qwen3-VL",
"section_id": "agent-docs/models/qwen3.md:L61",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/agent-docs/models/qwen3.md... | halo-023f0e6d0518f7e5f839 |
[
{
"role": "user",
"content": "What data format does Halo Online GRPO (RLVR) require?"
},
{
"role": "assistant",
"content": "Two columns: the prompt, and the ground truth the grader checks against.\n\n```jsonl\n{\"prompt\": \"What is 2 + 3? Put your answer in \\\\boxed{}.\", \"answer\": \"5\"}\n{... | concepts-and-behavior | [
{
"path": "human-docs/training-methods/online-grpo.md",
"section": "Online GRPO (RLVR) / Data",
"section_id": "human-docs/training-methods/online-grpo.md:L18",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff... | halo-027083818dddbea02ce5 |
[
{
"role": "user",
"content": "How do I train all weights with EP4 for Halo / Qwen3 MoE cookbook?"
},
{
"role": "assistant",
"content": "Create `qwen3-moe-sft.yaml`.\n\n```yaml\nmodel_name_or_path: Qwen/Qwen3-30B-A3B-Instruct-2507\nmodel_revision: 0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe\nmodel_i... | commands-and-recipes | [
{
"path": "human-docs/cookbooks/halo-qwen3-moe-cookbook.md",
"section": "Halo / Qwen3 MoE cookbook / Train all weights with EP4",
"section_id": "human-docs/cookbooks/halo-qwen3-moe-cookbook.md:L56",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo... | halo-028c2b5ea1fbb0dc0c65 |
[
{
"role": "user",
"content": "In Halo Training Methods / Methods at a glance, what are the details for Reward modeling?"
},
{
"role": "assistant",
"content": "| Method | Trainer | Dataset columns | When to use |\n|---|---|---|---|\n| Reward modeling | `DistributedRewardTrainer` | `chosen`, `reje... | concepts-and-behavior | [
{
"path": "agent-docs/training-methods/README.md",
"section": "Training Methods / Methods at a glance",
"section_id": "agent-docs/training-methods/README.md:L5",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-03173b70e21f40900fc0 |
[
{
"role": "user",
"content": "While using Halo, I see this symptom: **Async GRPO with Environments: completions fill with `Error: <tool>: missing a required argument: 'code'`**, episodes loop to `max_turns`, `episode/length_cutoff_turns` stays near zero. What causes it and how can I fix it?"
},
{
"r... | troubleshooting | [
{
"path": "agent-docs/reference/troubleshooting.md",
"section": "Troubleshooting / Symptom → cause → fix",
"section_id": "agent-docs/reference/troubleshooting.md:L22",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac31... | halo-0317671c8164ee8ea63e |
[
{
"role": "user",
"content": "For Halo Async GRPO with Environments, what should I know about The environments?"
},
{
"role": "assistant",
"content": "`environment_type` picks one; the last column says whether it grades against your dataset's `answer` column.\n\n| `environment_type` | The task |... | concepts-and-behavior | [
{
"path": "human-docs/training-methods/async-grpo-environments.md",
"section": "Async GRPO with Environments / The environments",
"section_id": "human-docs/training-methods/async-grpo-environments.md:L20",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecirc... | halo-03d5451b300529e41654 |
[
{
"role": "user",
"content": "In Halo Self-Distillation / Configuration, what does `reference_kl_coef` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "The config class is TRL's `SFTConfig`, so `max_length` defaults to `1024` — set it explicitly.\n\n```yaml\... | configuration | [
{
"path": "agent-docs/training-methods/distillation/self-distillation.md",
"section": "Self-Distillation / Configuration",
"section_id": "agent-docs/training-methods/distillation/self-distillation.md:L11",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecirc... | halo-03d93ad747e493a6db7e |
[
{
"role": "user",
"content": "What data format does Halo Offline GRPO require?"
},
{
"role": "assistant",
"content": "Three columns. **A group is one row**: put every completion of a prompt in the same row, because two rows sharing a\nprompt are two groups and are normalized separately.\n\n| Col... | concepts-and-behavior | [
{
"path": "human-docs/training-methods/offline-grpo.md",
"section": "Offline GRPO / Data",
"section_id": "human-docs/training-methods/offline-grpo.md:L16",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227d... | halo-03fb2a1484ccb0dddc1e |
[
{
"role": "user",
"content": "For Halo Tensor Parallelism (TP), what should I know about Multi-node?"
},
{
"role": "assistant",
"content": "TP requires an NVLink-bandwidth all-reduce after every layer, so it must stay **within a domain**;\ncross-node TP over InfiniBand would bottleneck every pas... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/tensor-parallelism.md",
"section": "Tensor Parallelism (TP) / Multi-node",
"section_id": "agent-docs/parallelism/tensor-parallelism.md:L283",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce75... | halo-04186886fd88e995697a |
[
{
"role": "user",
"content": "In Halo Data Parallelism / Limitations, what does `use_hsdp` on a single NVLink domain control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "**Trainers.** All of them — plain DP is the fallback every trainer runs on, and no trainer d... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/data-parallelism.md",
"section": "Data Parallelism / Limitations",
"section_id": "agent-docs/parallelism/data-parallelism.md:L243",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146... | halo-041fe6418f6df2b949df |
[
{
"role": "user",
"content": "While using Halo, I see this symptom: Every GPU at **100% utilization but idle power**, all ranks spinning a CPU core, no NCCL error. What causes it and how can I fix it?"
},
{
"role": "assistant",
"content": "| Symptom | Cause | Fix | Deep page |\n|---|---|---|---|... | troubleshooting | [
{
"path": "agent-docs/reference/troubleshooting.md",
"section": "Troubleshooting / Symptom → cause → fix",
"section_id": "agent-docs/reference/troubleshooting.md:L22",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac31... | halo-042de99942ba6ed81f45 |
[
{
"role": "user",
"content": "In Halo Configuration Reference / OfflineGRPOConfig, what does `policy_gradient_formulation` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "`OfflineGRPOTrainer`: group-relative policy optimization on pre-computed rewards. **So... | configuration | [
{
"path": "agent-docs/reference/configuration-reference.md",
"section": "Configuration Reference / OfflineGRPOConfig",
"section_id": "agent-docs/reference/configuration-reference.md:L346",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e... | halo-047f9a0283a0e7dd06ab |
[
{
"role": "user",
"content": "In Halo SFT Dataset Pre-Processing / Parameters, what does `--private` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "| Parameter | Default | Description |\n|-----------|---------|-------------|\n| `--private` | `False` | Make... | configuration | [
{
"path": "agent-docs/data/dataset-preparation.md",
"section": "SFT Dataset Pre-Processing / Parameters",
"section_id": "agent-docs/data/dataset-preparation.md:L117",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac314... | halo-048c5d80925a0c102cd9 |
[
{
"role": "user",
"content": "In Halo Expert Parallelism (EP) / Supported models, what are the details for GLM-5 Next?"
},
{
"role": "assistant",
"content": "| Model | HF Class | EP Wrapper | Per-family page |\n|---|---|---|---|\n| GLM-5 Next | `Glm5NextTextMoE` | `EPGlm5NextMoELayer` | glm5-nex... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/expert-parallelism.md",
"section": "Expert Parallelism (EP) / Supported models",
"section_id": "agent-docs/parallelism/expert-parallelism.md:L19",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963... | halo-0492cbfc20b831921230 |
[
{
"role": "user",
"content": "In Halo Configuration Guide / Parallelism parameters, what does `--ep_lazy_loading` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\ns... | configuration | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-049585d830444297fe36 |
[
{
"role": "user",
"content": "In Halo Configuration Reference / AsyncTrainingConfig, what does `advantage_quantile` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "`DistributedAsyncEnvironmentalGRPOTrainer`: Ray workers, rollout-server connections, weight s... | configuration | [
{
"path": "agent-docs/reference/configuration-reference.md",
"section": "Configuration Reference / AsyncTrainingConfig",
"section_id": "agent-docs/reference/configuration-reference.md:L494",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c... | halo-0522b195b3c1dad2de0f |
[
{
"role": "user",
"content": "Can you explain how the trainer finds it for Halo Rollout Servers?"
},
{
"role": "assistant",
"content": "| Config | Method | Meaning |\n| --- | --- | --- |\n| `rollout_server_url` | async GRPO | one server, default `http://localhost:8000` |\n| `rollout_server_confi... | concepts-and-behavior | [
{
"path": "human-docs/rollout-servers.md",
"section": "Rollout Servers / How the trainer finds it",
"section_id": "human-docs/rollout-servers.md:L100",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/... | halo-052d5ac776fd1fe22cdf |
[
{
"role": "user",
"content": "In Halo Docker Guide / Image matrix, what are the details for EFA userspace — AWS libfabric + `aws-ofi-nccl` at one pinned build (`docker/efa/install_efa_userspace.sh`), shared with the vLLM/SGLang images?"
},
{
"role": "assistant",
"content": "| Component | Hopper ... | concepts-and-behavior | [
{
"path": "agent-docs/infrastructure/docker.md",
"section": "Docker Guide / Image matrix",
"section_id": "agent-docs/infrastructure/docker.md:L67",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/agen... | halo-0530edec48660b038854 |
[
{
"role": "user",
"content": "In Halo AI Tooling / Using it, what are the details for `add-model`?"
},
{
"role": "assistant",
"content": "From a shell inside the container (bring your own subscription or API key):\n\n```bash\ncd /workspace && claude\n```\n\nDescribe the task. Four skills — `para... | concepts-and-behavior | [
{
"path": "human-docs/ai-tooling.md",
"section": "AI Tooling / Using it",
"section_id": "human-docs/ai-tooling.md:L9",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/human-docs/ai-tooling.md#L10-L33"... | halo-0541e019fcad9d262fd5 |
[
{
"role": "user",
"content": "In Halo Expert Parallelism (EP) / Supported models, what are the details for Mistral4?"
},
{
"role": "assistant",
"content": "| Model | HF Class | EP Wrapper | Per-family page |\n|---|---|---|---|\n| Mistral4 | `Mistral4MoE` | `EPMistral4MoELayer` | mistral4.md |\n\... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/expert-parallelism.md",
"section": "Expert Parallelism (EP) / Supported models",
"section_id": "agent-docs/parallelism/expert-parallelism.md:L19",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963... | halo-0567822b713354b29699 |
[
{
"role": "user",
"content": "For Halo GPT-OSS, what should I know about Serving for GRPO (vLLM)?"
},
{
"role": "assistant",
"content": "Use the toolkit `vllm-server:0.26.0` image, not stock upstream vLLM: on B300 the stock `vllm/vllm-openai` image produces garbage GPT-OSS output with harmony on... | concepts-and-behavior | [
{
"path": "agent-docs/models/gpt-oss.md",
"section": "GPT-OSS / Serving for GRPO (vLLM)",
"section_id": "agent-docs/models/gpt-oss.md:L100",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/agent-docs/... | halo-057b9cbdc1696ab8cf79 |
[
{
"role": "user",
"content": "Can you explain why bf16 wins for Halo Mixed-Precision Training: FP8 / FP4 (Low-Precision Compute)?"
},
{
"role": "assistant",
"content": "At production MoE shapes — EP degree 8, 256–512 tokens per local expert, expert width N ≤ 4096 (gpt-oss-120B N=2880, qwen3.6 do... | internals | [
{
"path": "agent-docs/optimization/low-precision-moe-kernels.md",
"section": "Mixed-Precision Training: FP8 / FP4 (Low-Precision Compute) / Why bf16 wins",
"section_id": "agent-docs/optimization/low-precision-moe-kernels.md:L23",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "http... | halo-05db8e02218d715c33b8 |
[
{
"role": "user",
"content": "For Halo Reward Terms, what should I know about Generative judge?"
},
{
"role": "assistant",
"content": "An OpenAI-compatible chat model reads the task, the response and the rubric, and replies with one JSON object of integer scores, one per requirement, from 0 to `... | concepts-and-behavior | [
{
"path": "agent-docs/training-methods/grpo/rewards.md",
"section": "Reward Terms / Generative judge",
"section_id": "agent-docs/training-methods/grpo/rewards.md:L30",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac31... | halo-0657ef6a7446bb4d1e60 |
[
{
"role": "user",
"content": "In Halo Rollout Servers / vLLM / Config-schema parity, what does `VLLM_CUDA_DEVICES` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "`Dockerfile.vllm` builds `vllm-server:0.26.0` with the native NCCL weight-transfer engine, the... | configuration | [
{
"path": "agent-docs/infrastructure/rollout-servers.md",
"section": "Rollout Servers / vLLM",
"section_id": "agent-docs/infrastructure/rollout-servers.md:L292",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-06619ae03baf85efe66d |
[
{
"role": "user",
"content": "In Halo Why this framework / Framework comparison, what are the details for Multi-node?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\nsurface, the rank math, the trainer gates, an... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-066b83894885c7ca4d9c |
[
{
"role": "user",
"content": "What should I know about Halo Rollout Servers? Focus on practical usage."
},
{
"role": "assistant",
"content": "On-policy RL needs the model to generate while it trains. Halo does that in a **separate inference container** —\nvLLM or SGLang — that the trainer talks ... | concepts-and-behavior | [
{
"path": "human-docs/rollout-servers.md",
"section": "Rollout Servers",
"section_id": "human-docs/rollout-servers.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/human-docs/rollout-servers.md... | halo-067f155797693107a7dd |
[
{
"role": "user",
"content": "How do I train Laguna XS 2.1 on one GPU for Halo / Poolside Laguna 2.1 cookbook?"
},
{
"role": "assistant",
"content": "Copy the config to `laguna-xs-2.1-sft.yaml`, change the model and output directory, and\ndrop `expert_parallel_size`.\n\n```yaml\nmodel_name_or_pa... | commands-and-recipes | [
{
"path": "human-docs/cookbooks/halo-laguna-2.1-cookbook.md",
"section": "Halo / Poolside Laguna 2.1 cookbook / Train Laguna XS 2.1 on one GPU",
"section_id": "human-docs/cookbooks/halo-laguna-2.1-cookbook.md:L140",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com... | halo-06a8b9294d00d3cafa3b |
[
{
"role": "user",
"content": "While using Halo, I see this symptom: **`EP capacity dedup: a MoE layer dispatched N tokens/rank, over the capacity C cached at forward generation G`** on one rank, its peers stuck in the dispatch. What causes it and how can I fix it?"
},
{
"role": "assistant",
"con... | troubleshooting | [
{
"path": "agent-docs/reference/troubleshooting.md",
"section": "Troubleshooting / Symptom → cause → fix",
"section_id": "agent-docs/reference/troubleshooting.md:L22",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac31... | halo-06ab3137b5051e4576dc |
[
{
"role": "user",
"content": "While using Halo, I see this symptom: Nodes never join, or the rendezvous times out. What causes it and how can I fix it?"
},
{
"role": "assistant",
"content": "| Symptom | Cause → fix |\n| --- | --- |\n| Nodes never join, or the rendezvous times out | The torchrun ... | troubleshooting | [
{
"path": "human-docs/troubleshooting.md",
"section": "Troubleshooting / Multi-node and clusters",
"section_id": "human-docs/troubleshooting.md:L33",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/hu... | halo-06e5bbdf40cf4b726b18 |
[
{
"role": "user",
"content": "For Halo Architecture, what should I know about Trainers?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\nsurface, the rank math, the trainer gates, and the stage/split/loss/checkpo... | internals | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-06f0247c973d3449ab5e |
[
{
"role": "user",
"content": "Can you explain what every method shares for Halo Training Methods?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\nsurface, the rank math, the trainer gates, and the stage/split/lo... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-06f96f6598c16feff957 |
[
{
"role": "user",
"content": "In Halo Offline GRPO / Configuration, what does `best_completion_emphasis` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "```yaml\nmodel_name_or_path: Qwen/Qwen3.6-35B-A3B\ndataset:\n - s3://bucket/gsm8k-qwen3.6-35b-a3b-offli... | configuration | [
{
"path": "agent-docs/training-methods/grpo/offline-grpo.md",
"section": "Offline GRPO / Configuration",
"section_id": "agent-docs/training-methods/grpo/offline-grpo.md:L31",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce752... | halo-06fe38b8fad27dc4e5d0 |
[
{
"role": "user",
"content": "In Halo Metrics and Troubleshooting / Logged metrics, what are the details for unnamespaced?"
},
{
"role": "assistant",
"content": "| Namespace | Answers | Keys to watch |\n|---|---|---|\n| unnamespaced | TRL loss internals | `clip_ratio/*`, `cispo_clip_ratio`, `ste... | troubleshooting | [
{
"path": "agent-docs/training-methods/grpo/async-grpo/monitoring.md",
"section": "Metrics and Troubleshooting / Logged metrics",
"section_id": "agent-docs/training-methods/grpo/async-grpo/monitoring.md:L5",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whiteci... | halo-071805833fc172e0c4c8 |
[
{
"role": "user",
"content": "For Halo Parallelism, what should I know about torch.compile?"
},
{
"role": "assistant",
"content": "On EP MoE, `torch_compile` reaches about the same speedup as Liger kernels (the default) and composes with them. DeepEP all-to-all (EP), DTensor dispatch (TP), and U... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/README.md",
"section": "Parallelism / torch.compile",
"section_id": "agent-docs/parallelism/README.md:L131",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/agent-doc... | halo-07371635d4b26326cfee |
[
{
"role": "user",
"content": "In Halo Why this framework / Framework comparison / When to use which, what are the details for EP+TP or pure ETP on MoE?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\nsurface, th... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-0760d670459ae2d0562f |
[
{
"role": "user",
"content": "What should I know about Halo Infrastructure?"
},
{
"role": "assistant",
"content": "Runtime image, cloud deployments, and hardware-specific dependency installation."
}
] | concepts-and-behavior | [
{
"path": "agent-docs/infrastructure/README.md",
"section": "Infrastructure",
"section_id": "agent-docs/infrastructure/README.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/agent-docs/infrast... | halo-07878b5f6bcbb06c3c9b |
[
{
"role": "user",
"content": "Which seemingly useful optimizations can fail to help Halo MoE training, and why?"
},
{
"role": "assistant",
"content": "| Lever | Why not |\n| --- | --- |\n| fp8 / fp4 **compute** (`lowp_precision`) | fine-grained experts are weight-bandwidth-bound, not FLOP-bound,... | concepts-and-behavior | [
{
"path": "human-docs/performance.md",
"section": "Performance / Levers that do not help here",
"section_id": "human-docs/performance.md:L48",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/human-doc... | halo-079def186378f218408f |
[
{
"role": "user",
"content": "For Halo Step-3.7 Flash, what should I know about Serving for GRPO (vLLM)?"
},
{
"role": "assistant",
"content": "Online and async GRPO sync into vLLM 0.26.0, which serves the family from the **hub** namespace (`model.layers.N.moe.gate_proj` / `up_proj` / `down_proj... | concepts-and-behavior | [
{
"path": "agent-docs/models/step3p7.md",
"section": "Step-3.7 Flash / Serving for GRPO (vLLM)",
"section_id": "agent-docs/models/step3p7.md:L74",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/agent... | halo-07a76f0af47aa28e0362 |
[
{
"role": "user",
"content": "In Halo Nomad Deployment / Mapping onto `docker run`, what are the details for `--ipc=host`?"
},
{
"role": "assistant",
"content": "The specs carry the same container flags every other Halo surface passes (`CLAUDE.md`, the Makefile, the SkyPilot tasks):\n\n| `docker... | commands-and-recipes | [
{
"path": "agent-docs/infrastructure/nomad.md",
"section": "Nomad Deployment / Mapping onto `docker run`",
"section_id": "agent-docs/infrastructure/nomad.md:L168",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146cc... | halo-07e51511168217bd0e6d |
[
{
"role": "user",
"content": "For Halo Tensor Parallelism (TP) / The selective-TP plan, what should I know about Identical-input requirement?"
},
{
"role": "assistant",
"content": "`apply_tp_to_attention_only` (`src/distributed/tensor_parallel/parallelize_attention.py`) walks the\nmodel and buil... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/tensor-parallelism.md",
"section": "Tensor Parallelism (TP) / The selective-TP plan",
"section_id": "agent-docs/parallelism/tensor-parallelism.md:L113",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c... | halo-07f992f5d0c7312d73ca |
[
{
"role": "user",
"content": "In Halo Rollout Servers / SGLang, what does `SGLANG_PORT` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "`Dockerfile.sglang` builds `sglang-server:0.5.17` to align NCCL: upstream's wheel trails `uv.lock`'s\nexact pin (what the... | concepts-and-behavior | [
{
"path": "agent-docs/infrastructure/rollout-servers.md",
"section": "Rollout Servers / SGLang",
"section_id": "agent-docs/infrastructure/rollout-servers.md:L451",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146cc... | halo-082f0679eefe47d7a2db |
[
{
"role": "user",
"content": "In Halo Environment Variables / Tuning knobs worth knowing, what does `HALO_DEEPEP_GPU_TIMEOUT_SECONDS` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "Halo has around twenty more `HALO_*` knobs, all optional and all defaulted ... | configuration | [
{
"path": "human-docs/environment-variables.md",
"section": "Environment Variables / Tuning knobs worth knowing",
"section_id": "human-docs/environment-variables.md:L73",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bca... | halo-083ae9c18ce59a919598 |
[
{
"role": "user",
"content": "What should I know about Halo / Command A+ cookbook?"
},
{
"role": "assistant",
"content": "Fine-tune Cohere Command A+ with Halo.\n\nThis recipe uses the BF16 checkpoint and text-only UltraChat data. The same model can also train on image-text data.\n\nCommand A+ h... | commands-and-recipes | [
{
"path": "human-docs/cookbooks/halo-command-a-plus-cookbook.md",
"section": "Halo / Command A+ cookbook",
"section_id": "human-docs/cookbooks/halo-command-a-plus-cookbook.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963d... | halo-0880b3d316e78f36d19e |
[
{
"role": "user",
"content": "In Halo Configuration Reference / Sequence length: caps vs generation budgets / Per-trainer mechanism, what are the details for SMPO (`SmoothMarginPOConfig`)?"
},
{
"role": "assistant",
"content": "Three knob names carry the length settings. They mean the same thing... | configuration | [
{
"path": "agent-docs/reference/configuration-reference.md",
"section": "Configuration Reference / Sequence length: caps vs generation budgets",
"section_id": "agent-docs/reference/configuration-reference.md:L14",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/w... | halo-08acd685056fd6cfd58f |
[
{
"role": "user",
"content": "In Halo Docker Guide / Baked-in environment variables, what does `CUTE_DSL_ENABLE_TVM_FFI` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "| Variable | Value | Effect |\n|----------|-------|--------|\n| `CUTE_DSL_ENABLE_TVM_FFI... | configuration | [
{
"path": "agent-docs/infrastructure/docker.md",
"section": "Docker Guide / Baked-in environment variables",
"section_id": "agent-docs/infrastructure/docker.md:L160",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac314... | halo-08ad4b6c828f95dde838 |
[
{
"role": "user",
"content": "While using Halo, I see this symptom: `gradient_checkpointing_enable` raises on Zaya. What causes it and how can I fix it?"
},
{
"role": "assistant",
"content": "| Symptom | Cause | Fix | Deep page |\n|---|---|---|---|\n| `gradient_checkpointing_enable` raises on Za... | troubleshooting | [
{
"path": "agent-docs/reference/troubleshooting.md",
"section": "Troubleshooting / Symptom → cause → fix",
"section_id": "agent-docs/reference/troubleshooting.md:L22",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac31... | halo-08b1a78e7a4dc9cd1d10 |
[
{
"role": "user",
"content": "In Halo Configuration Reference / SmoothMarginPOConfig, what does `model_init_kwargs` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "`SmoothMarginPOTrainer` (SMPO) — reference-model-free, margin-based preference optimization. ... | configuration | [
{
"path": "agent-docs/reference/configuration-reference.md",
"section": "Configuration Reference / SmoothMarginPOConfig",
"section_id": "agent-docs/reference/configuration-reference.md:L371",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4... | halo-08b54f29c01c9cb164eb |
[
{
"role": "user",
"content": "For Halo Preference Tuning, what should I know about Length budgets?"
},
{
"role": "assistant",
"content": "SMPO splits `max_length` into a prompt share and a completion share: `max_prompt_length` defaults to\nhalf of it and the completion takes the remainder, promp... | concepts-and-behavior | [
{
"path": "human-docs/training-methods/preference.md",
"section": "Preference Tuning / Length budgets",
"section_id": "human-docs/training-methods/preference.md:L110",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac31... | halo-08b7b97a1b90b2decaa7 |
[
{
"role": "user",
"content": "For Halo GPT-OSS, what should I know about Attention sinks?"
},
{
"role": "assistant",
"content": "GptOss attention adds a per-head learnable scalar to the attention logits before softmax, with two parallelism consequences.\n\n**TP sharding.** `shard_sinks_param()` ... | concepts-and-behavior | [
{
"path": "agent-docs/models/gpt-oss.md",
"section": "GPT-OSS / Attention sinks",
"section_id": "agent-docs/models/gpt-oss.md:L35",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/agent-docs/models/gp... | halo-08c5974abdefb321297d |
[
{
"role": "user",
"content": "What dataset structure does Halo Classification require?"
},
{
"role": "assistant",
"content": "A `prompt` conversation plus a `label` column: a string for single-label, a list for multi-label. `text_field` names a raw text column instead, wrapped as one user turn. ... | concepts-and-behavior | [
{
"path": "agent-docs/training-methods/classification.md",
"section": "Classification / Dataset",
"section_id": "agent-docs/training-methods/classification.md:L7",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146cc... | halo-08d3021715c5252fe82f |
[
{
"role": "user",
"content": "How do I train all weights with EP8 for Fine-tune Gemma 4 MoE with Halo?"
},
{
"role": "assistant",
"content": "Create `gemma4-sft.yaml`, or start from the shipped\n`examples/sft/gemma4/gemma4-26b-a4b-ultrachat-ep.yaml`.\n\n```yaml\nmodel_name_or_path: google/gemma-... | commands-and-recipes | [
{
"path": "human-docs/cookbooks/halo-gemma4-moe-cookbook.md",
"section": "Fine-tune Gemma 4 MoE with Halo / Train all weights with EP8",
"section_id": "human-docs/cookbooks/halo-gemma4-moe-cookbook.md:L58",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecir... | halo-08de8ad1141de87c175f |
[
{
"role": "user",
"content": "For Halo Clusters and Multi-Node, what should I know about SkyPilot?"
},
{
"role": "assistant",
"content": "`launcher-configs/skypilot/` holds launchable task YAMLs for AWS and Nebius:\nGPT-OSS 20B (node-local EP), plus GPT-OSS 120B and Qwen3.5-122B in node-local\na... | concepts-and-behavior | [
{
"path": "human-docs/clusters.md",
"section": "Clusters and Multi-Node / SkyPilot",
"section_id": "human-docs/clusters.md:L112",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/human-docs/clusters.md... | halo-08e0a02ad440c61274ed |
[
{
"role": "user",
"content": "For Halo Collators, what should I know about Document isolation under packing?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\nsurface, the rank math, the trainer gates, and the sta... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-0905d644a0d7db39ab69 |
[
{
"role": "user",
"content": "In Halo Offline GRPO / Configuration, what does `loss_type` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "```yaml\nmodel_name_or_path: Qwen/Qwen3.6-35B-A3B\ndataset:\n - s3://bucket/gsm8k-qwen3.6-35b-a3b-offline-grpo\nadvant... | configuration | [
{
"path": "agent-docs/training-methods/grpo/offline-grpo.md",
"section": "Offline GRPO / Configuration",
"section_id": "agent-docs/training-methods/grpo/offline-grpo.md:L31",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce752... | halo-09291e4b0cb172a6a8f1 |
[
{
"role": "user",
"content": "In Halo Glossary / Parallelism, what are the details for **Expert Tensor Parallel (ETP)**?"
},
{
"role": "assistant",
"content": "| Term | Definition |\n|---|---|\n| **Expert Tensor Parallel (ETP)** | Shards each expert's FFN weights across `expert_tp_size` GPUs — i... | concepts-and-behavior | [
{
"path": "agent-docs/reference/glossary.md",
"section": "Glossary / Parallelism",
"section_id": "agent-docs/reference/glossary.md:L9",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/agent-docs/refer... | halo-09431ce4120ca0171f71 |
[
{
"role": "user",
"content": "For Halo vs stock TRL / transformers, what should I know about Throughput & memory (4k–16k)?"
},
{
"role": "assistant",
"content": "*GC-on · grouped GEMM · TRL ZeRO-3 · Halo ZeRO-2 (EP1 also at ZeRO-3) · elastic.* **tok/s/GPU · peak GB:**\n\n| seq·b | stock TRL (z3)... | concepts-and-behavior | [
{
"path": "agent-docs/optimization/halo-vs-stock-trl.md",
"section": "Halo vs stock TRL / transformers / Throughput & memory (4k–16k)",
"section_id": "agent-docs/optimization/halo-vs-stock-trl.md:L26",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/h... | halo-094ade7ee14d88c63f43 |
[
{
"role": "user",
"content": "In Halo Troubleshooting / Parallelism config rejections, what are the details for `EP + TP + ETP`?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\nsurface, the rank math, the traine... | troubleshooting | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-0959324d70009adb770c |
[
{
"role": "user",
"content": "In Halo Supported Matrix / Training methods, what are the details for Embedding?"
},
{
"role": "assistant",
"content": "Every trainer supports EP, TP, ETP and EP+TP. CP is declare-to-enable and only\nSFT and SMPO set it: nothing inspects a trainer's loss, so CP woul... | concepts-and-behavior | [
{
"path": "human-docs/supported-matrix.md",
"section": "Supported Matrix / Training methods",
"section_id": "human-docs/supported-matrix.md:L40",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/human-... | halo-09597d64d20e17cccfa7 |
[
{
"role": "user",
"content": "In Halo Collators / Document isolation under packing, what are the details for Qwen3.5 / 3.6, Qwen3-Next?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\nsurface, the rank math, the... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-095e22ad3d6b44652986 |
[
{
"role": "user",
"content": "In Halo Rollout Servers / vLLM / Config-schema parity, what does `VLLM_MOE_BACKEND` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "`Dockerfile.vllm` builds `vllm-server:0.26.0` with the native NCCL weight-transfer engine, the\... | configuration | [
{
"path": "agent-docs/infrastructure/rollout-servers.md",
"section": "Rollout Servers / vLLM",
"section_id": "agent-docs/infrastructure/rollout-servers.md:L292",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-096548d5fa8c33c1d904 |
[
{
"role": "user",
"content": "Can you explain why CP and TP are unavailable for Halo Laguna?"
},
{
"role": "assistant",
"content": "Both are registry-gated, and Laguna is in neither registry:\n\n- **CP** needs a Ulysses attention wrapper. `LagunaAttention` is absent from `CP_SUPPORTED_ATTENTION_... | concepts-and-behavior | [
{
"path": "agent-docs/models/laguna.md",
"section": "Laguna / Why CP and TP are unavailable",
"section_id": "agent-docs/models/laguna.md:L54",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/agent-doc... | halo-097a11bac7895680f02e |
[
{
"role": "user",
"content": "What should I know about Halo GPU Training Theory?"
},
{
"role": "assistant",
"content": "Every \"should I use X?\" reduces to one question: which bottleneck does X attack, and is that the one you have? The **roofline** (§2) answers it; §3 names the four bottlenecks... | concepts-and-behavior | [
{
"path": "agent-docs/reference/gpu-training-theory.md",
"section": "GPU Training Theory",
"section_id": "agent-docs/reference/gpu-training-theory.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc... | halo-09a98d769de72b7a8029 |
[
{
"role": "user",
"content": "How do I launch Halo Teacher Distillation?"
},
{
"role": "assistant",
"content": "```bash\ntorchrun --nproc_per_node=8 scripts/training/distillation/teacher_distill.py \\\n examples/distillation/qwen3_5/distill-qwen3.5-9b-from-qwen3.6-35b-a3b.yaml\n```\n\n`halo l... | commands-and-recipes | [
{
"path": "agent-docs/training-methods/distillation/teacher-distillation.md",
"section": "Teacher Distillation / Launch",
"section_id": "agent-docs/training-methods/distillation/teacher-distillation.md:L64",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whiteci... | halo-09b516250c948a473406 |
[
{
"role": "user",
"content": "In Halo Supported Models / Choosing a parallelism configuration, what are the details for Same MoEs, long context?"
},
{
"role": "assistant",
"content": "| You have | Sequence length | Recommended mode |\n|---|---|---|\n| Same MoEs, long context | > 32K | EP+CP |\n\... | configuration | [
{
"path": "agent-docs/models/README.md",
"section": "Supported Models / Choosing a parallelism configuration",
"section_id": "agent-docs/models/README.md:L117",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaff... | halo-0a1a5ea60dd8b5c612a5 |
[
{
"role": "user",
"content": "In Halo Scripts Reference / Preparation scripts, what are the details for `scripts/before_training/convert_mistral4_bf16.py`?"
},
{
"role": "assistant",
"content": "| Script | Description |\n|--------|-------------|\n| `scripts/before_training/convert_mistral4_bf16.... | concepts-and-behavior | [
{
"path": "agent-docs/reference/scripts-reference.md",
"section": "Scripts Reference / Preparation scripts",
"section_id": "agent-docs/reference/scripts-reference.md:L289",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520b... | halo-0a24b96952b164e855bc |
[
{
"role": "user",
"content": "For Halo Mistral4 (text backbone of Mistral3 VLMs), what should I know about Serving?"
},
{
"role": "assistant",
"content": "vLLM 0.26.0 knows nothing of this family: `mistral4` appears in neither its architecture registry nor any `model_type` mapping. The composite... | concepts-and-behavior | [
{
"path": "agent-docs/models/mistral4.md",
"section": "Mistral4 (text backbone of Mistral3 VLMs) / Serving",
"section_id": "agent-docs/models/mistral4.md:L99",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff... | halo-0a4ecabb2980fb3b5eca |
[
{
"role": "user",
"content": "In Halo Configuration Reference / ClassificationConfig, what does `disable_dropout` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "`ClassificationTrainer`. **Source:** `src/configs/classification_config.py`\n\n| Parameter | Ty... | configuration | [
{
"path": "agent-docs/reference/configuration-reference.md",
"section": "Configuration Reference / ClassificationConfig",
"section_id": "agent-docs/reference/configuration-reference.md:L432",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4... | halo-0a545e8e5d258c39062d |
[
{
"role": "user",
"content": "In Halo Model Merging / Usage, what does `--models` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "```bash\n# Weighted linear average (weights are normalized by their sum)\npython scripts/after_training/merge_models.py --metho... | commands-and-recipes | [
{
"path": "agent-docs/reference/model-merging.md",
"section": "Model Merging / Usage",
"section_id": "agent-docs/reference/model-merging.md:L16",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/agent-... | halo-0a548355cf3cd81d4319 |
[
{
"role": "user",
"content": "What should I know about Halo Mistral4 (text backbone of Mistral3 VLMs)?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\nsurface, the rank math, the trainer gates, and the stage/spl... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-0a63d628d7208dff6a32 |
[
{
"role": "user",
"content": "In Halo Configuration Reference / ParallelismConfig, what does `lowp_apply_moe_experts` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the confi... | configuration | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-0a770fcc0a308bfa16d5 |
[
{
"role": "user",
"content": "In Halo Glossary / Methods and infrastructure, what are the details for **R3 (routing replay)**?"
},
{
"role": "assistant",
"content": "| Term | Definition |\n|---|---|\n| **R3 (routing replay)** | MoE rollout↔training routing alignment: the rollout engine records w... | concepts-and-behavior | [
{
"path": "agent-docs/reference/glossary.md",
"section": "Glossary / Methods and infrastructure",
"section_id": "agent-docs/reference/glossary.md:L60",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/... | halo-0aa9ea0438fb44b4a97e |
[
{
"role": "user",
"content": "In Halo Configuration Reference / Script arguments / CommonScriptArguments, what does `pad_token` / `bos_token` / `eos_token` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "Base shared by all training scripts. **Source:** `src... | configuration | [
{
"path": "agent-docs/reference/configuration-reference.md",
"section": "Configuration Reference / Script arguments / CommonScriptArguments",
"section_id": "agent-docs/reference/configuration-reference.md:L548",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whi... | halo-0ac591e037b1849ac5f2 |
[
{
"role": "user",
"content": "In Halo Supported Matrix / Model families, what are the details for Any other HF causal LM?"
},
{
"role": "assistant",
"content": "Any HuggingFace `AutoModelForCausalLM` runs under FSDP2; this table is about\nadvanced parallelism. Qwen3 MoE, GPT-OSS, GLM-4 MoE Lite ... | concepts-and-behavior | [
{
"path": "human-docs/supported-matrix.md",
"section": "Supported Matrix / Model families",
"section_id": "human-docs/supported-matrix.md:L99",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff227dc9f/human-do... | halo-0ad6da0048417e34da2e |
[
{
"role": "user",
"content": "What should I know about Halo Embedding Training?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\nsurface, the rank math, the trainer gates, and the stage/split/loss/checkpoint cont... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-0afb2d2ed369a5fd6fb5 |
[
{
"role": "user",
"content": "For Halo Configuration Reference, what should I know about EnvironmentConfig?"
},
{
"role": "assistant",
"content": "Environment selection, reward terms and turn budget for Async GRPO with Environments. **Source:** `src/configs/environment_config.py`\n\n| Parameter ... | configuration | [
{
"path": "agent-docs/reference/configuration-reference.md",
"section": "Configuration Reference / EnvironmentConfig",
"section_id": "agent-docs/reference/configuration-reference.md:L455",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e... | halo-0b1e0b4036fe62cdfe5f |
[
{
"role": "user",
"content": "In Halo Rollout Servers / SGLang, what does `SGLANG_ATTENTION_BACKEND` control and what settings or limitations apply?"
},
{
"role": "assistant",
"content": "`Dockerfile.sglang` builds `sglang-server:0.5.17` to align NCCL: upstream's wheel trails `uv.lock`'s\nexact ... | concepts-and-behavior | [
{
"path": "agent-docs/infrastructure/rollout-servers.md",
"section": "Rollout Servers / SGLang",
"section_id": "agent-docs/infrastructure/rollout-servers.md:L451",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146cc... | halo-0b29dd52ab6e6f5f42cf |
[
{
"role": "user",
"content": "In Halo Why this framework / Framework comparison, what are the details for Offline GRPO (pre-scored)?"
},
{
"role": "assistant",
"content": "**Status: not yet available in this release.** The pipeline-parallel seams ship — the config\nsurface, the rank math, the tr... | concepts-and-behavior | [
{
"path": "agent-docs/parallelism/pipeline-parallelism.md",
"section": "Pipeline Parallelism",
"section_id": "agent-docs/parallelism/pipeline-parallelism.md:L1",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccaf... | halo-0b2a5090165dcc499bbc |
[
{
"role": "user",
"content": "In Halo Quickstart / 2. Other methods, what are the details for DPO / KTO?"
},
{
"role": "assistant",
"content": "Same YAML shape, different script and a few method fields. `halo launch <method> <config>` runs any of them by name and picks the launcher (`halo launch... | commands-and-recipes | [
{
"path": "agent-docs/getting-started/quickstart.md",
"section": "Quickstart / 2. Other methods",
"section_id": "agent-docs/getting-started/quickstart.md:L47",
"commit": "4c1e6c65b963dce7520bcac3146ccafff227dc9f",
"url": "https://github.com/whitecircle/halo/blob/4c1e6c65b963dce7520bcac3146ccafff... | halo-0b57e8ef2522f18103a2 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.