laion/qwen3-30b-a3b-thinking-opencode-sft-densemixer-serveparity

# axolotl SFT — Run 1 (DenseMixer ON) — Qwen3-30B-A3B-Thinking-2507 on opencode traces.
# Experiment: axolotl-sft-opencode-densemoe (task #16). Paired-init controlled ablation:
#   Run 1 = densemixer ON (this file), Run 2 = byte-identical EXCEPT `dense_mixer: false`.
# Design/rationale: experiments/active/axolotl-sft-opencode-densemoe/{POLICY,STATE}.md.
#
# ⚠ LAUNCH PATH = DIRECT `axolotl.cli.train` (NOT hpc.launch). EmpireAI is NOT registered
# in hpc.launch; the SFT path is torch.distributed.run -m axolotl.cli.train <this.yaml>
# (bring-up Stage-3 recipe: ~/scripts/stage3_incontainer.sh, bond0 NCCL). The direct path
# honors this WHOLE file verbatim — which is REQUIRED here: the hpc.launch translator
# (hpc/axolotl_config_utils.py) would STRIP `plugins:`, `dense_mixer:` (rebuilds plugins
# from a fixed whitelist) AND `fp8:` (unhandled key) → densemixer silently off + fp8 auto-on.
#
# ⚠ REQUIRES the densemixer-enabled image `mega_final_dm.sqsh` (densemixer==1.0.1 in the SFT
# system python). The DenseMixerPlugin hard-raises "DenseMixer is not installed" otherwise.

# θ₀ — the SHARED init both runs start from (control discipline). Pinned local snapshot.
base_model: /mnt/home/bf996/experiments/densemixer/theta0    # Qwen/Qwen3-30B-A3B-Thinking-2507 @ 144afc2f...
model_type: AutoModelForCausalLM
trust_remote_code: true

# === THE DenseMoE mechanism = DenseMixer (marin #7088; yaof20/DenseMixer) ===
# Training-only patch of Qwen3MoeSparseMoeBlock.forward: dense forward through ALL experts +
# full-softmax router + STE combine (forward value = sparse/identical; backward = dense
# gradient) + per-expert grad hook (expert params still update sparsely). Net: the ROUTER
# gets the dense all-experts gradient. Inference untouched. This IS the single Run-1/Run-2
# differentiator — Run 2 flips ONLY `dense_mixer: false` (plugin becomes a no-op).
plugins:
  - axolotl.integrations.densemixer.DenseMixerPlugin
dense_mixer: true                 # Run 1 = ON.  Run 2 = false (only-flag diff).

# opencode SFT dataset — SERVE-PARITY REBUILD (bug ledger #2 fix). IDENTICAL for both runs.
# The old `conversations` field was LOSSY (no system, empty turn-0, no <tools>, tool calls as
# inline text, tool results as role:user) → model never learned the task/tools/read-edit workflow
# → bash-spammed → swebench ~0. Rebuilt from the literal token columns via
# scripts/harbor/literal_traces_to_opencode_sft.py: `messages` (list) carries a system turn (the
# opencode agent prompt), a task-grounded user turn-0, structured assistant `tool_calls` (NOT inline
# text, args JSON-string → axolotl json.loads → dict), and role:tool observations; `tools` (JSON
# string) is the 10 opencode function schemas. Source: open-athena/nemotron-code-oracle-opencode-sft-serveparity.
datasets:
  - path: /mnt/home/bf996/experiments/densemixer/data_opencode_serveparity
    ds_type: parquet
    data_files:
      - /mnt/home/bf996/experiments/densemixer/data_opencode_serveparity/data/train-*.parquet
    type: chat_template
    field_messages: messages
    field_tools: tools                    # render the <tools> system block (train==serve)
    message_property_mappings:
      role: role
      content: content
    split_thinking: false                 # keep <think> inline on all turns (Qwen3-Thinking parity)
# TRAIN==SERVE (POLICY §7): tokenizer_default = the base Qwen3-30B-A3B-Thinking tokenizer's own
# tools-aware template (4049 chars, `{% if tools %}` + `<tool_call>` + `<tool_response>`), byte-
# identical to what opencode serves at eval; archived as chat_templates/chat_template_v1_base-tools-aware.jinja.
# Also sidesteps axolotl export-strip bug #1: the saved template == the base tools-aware one (no bare-chatml clobber).
chat_template: tokenizer_default
dataset_prepared_path: /mnt/home/bf996/experiments/densemixer/prepared/run1_serveparity   # SHARED (multi-node sentinel); NEW path — old cache was the lossy tokenization
val_set_size: 0.0
dataset_num_proc: 1
# Dataloader first-batch fix — the default async path (num_workers:1 + prefetch_factor:256)
# wedged the multi-node first batch: the worker eagerly prefetched 256 packed 16384-tok batches
# by mmap-reading the NFS-backed prepared arrow -> rank-0 stuck in folio_wait (mmap page I/O) +
# an anon_pipe_read thread blocked on the worker, GPUs 0% (forward never started), no NCCL
# collective in flight (job 31682, killed ~21min pre-step). prefetch_factor:2 kills the storm.
dataloader_num_workers: 2
dataloader_prefetch_factor: 2

# === precision — bf16 + flash-attn (coordinator signal 2026-07-17; job 31623 HEALTHY). ===
bf16: true
fp16: false
fp8: false                        # ⚠ MANDATORY EXPLICIT — axolotl 0.17 auto-enables fp8 on sm_100 → nan.
tf32: false
attn_implementation: flash_attention_2   # load-bearing: FA2 avoids the Blackwell cuDNN-SDPA bf16 backward nan
                                          # (= the healthy job 31623 path; non-deprecated form of flash_attention:true)

# === memory: 30B MoE + DENSE forward (all 128 experts materialized ≈ 16× expert-FFN FLOPs). ===
# ZeRO-3 shards params/grads/optimizer; grad-checkpointing recomputes activations; chunked-CE
# bounds the LM-head logit tensor WITHOUT a model monkeypatch (avoids any liger×densemixer
# patch-interaction on qwen3_moe — deliberately NOT using enable_liger_kernel here).
deepspeed: /opt/axolotl/deepspeed_configs/zero3_bf16.json
gradient_checkpointing: true
chunked_cross_entropy: true
sequence_len: 16384               # conservative for the dense forward; raise to 32768 iff smoke shows headroom
sample_packing: true              # OPERATOR: packing ON (cadence)

# === control discipline — IDENTICAL both runs ===
seed: 42
micro_batch_size: 1
gradient_accumulation_steps: 4    # 8 GPUs (2×4) × micro 1 × accum 4 = effective batch 32
num_epochs: 3.0
learning_rate: 2.0e-5
lr_scheduler: cosine
warmup_ratio: 0.1
max_grad_norm: 1.0
optimizer: adamw_torch_fused
weight_decay: 0.0

# === checkpoint cadence (POLICY §5.5) — θ₀ + intermediate + final for the Δθ trajectory. ===
# save_steps → ~10-15 intermediate ckpts; FINALIZE once preprocess reveals the packed step-count.
# θ₀ (step 0) = the pinned prestaged snapshot. Keep ALL ckpts (Δθ analysis) — do NOT prune.
logging_steps: 1                  # instrument first 2-3 steps: grad_norm finite + loss real (nan → STOP→fp32)
save_steps: 10
save_total_limit: 100
output_dir: /mnt/home/bf996/experiments/densemixer/run1_dense_out
# Omit hub_model_id (offline init_hf_repo crash); push manually at cleanup. enable_db_registration: false.
special_tokens: {}