README.md

---
library_name: transformers
tags:
- generated_from_trainer
model-index:
- name: mnt/home/bf996/experiments/densemixer/run2_sparse_out
  results: []
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
<details><summary>See axolotl config</summary>

axolotl version: `0.17.0.dev0`
```yaml
# axolotl SFT — Run 2 (DenseMixer OFF = native sparse) — Qwen3-30B-A3B-Thinking-2507 on opencode traces.
# Experiment: axolotl-sft-opencode-densemoe (task #16). The CONTROL arm of the paired ablation:
#   BYTE-IDENTICAL to Run 1 (densemixer_run1_opencode.yaml) EXCEPT `dense_mixer: false` + `output_dir`.
# Control discipline (POLICY §2): SAME θ₀, seed, dataset + data-ORDER (same shared prepared path),
# seq_len, batch, packing, LR schedule, step count — the ONLY functional diff is dense_mixer true→false
# (plugin stays loaded; with false its pre_model_load is a no-op → stock sparse top-k MoE forward).
# Design/rationale: experiments/active/axolotl-sft-opencode-densemoe/{POLICY,STATE}.md.
#
# ⚠ LAUNCH PATH = DIRECT `axolotl.cli.train` (NOT hpc.launch — it would strip plugins/dense_mixer/fp8).
# ⚠ REQUIRES image `mega_final_dm.sqsh` (densemixer==1.0.1 + the tf-5.x port baked in).

# θ₀ — the SHARED init both runs start from (control discipline). IDENTICAL to Run 1.
base_model: /mnt/home/bf996/experiments/densemixer/theta0    # Qwen/Qwen3-30B-A3B-Thinking-2507 @ 144afc2f...
model_type: AutoModelForCausalLM
trust_remote_code: true

# === THE one-flag control diff: DenseMixer OFF ===
# Plugin stays in the stack (identical to Run 1); `dense_mixer: false` makes its pre_model_load a
# no-op → the model keeps the STOCK sparse top-k Qwen3MoE forward (non-selected experts' router
# logits get zero task-loss gradient). This is the sparse baseline for the Δθ counterfactual.
plugins:
  - axolotl.integrations.densemixer.DenseMixerPlugin
dense_mixer: false                # Run 2 = OFF (the ONLY functional diff vs Run 1).

# opencode SFT dataset — IDENTICAL pinned revision + SHARED prepared path (guarantees same data ORDER).
datasets:
  - path: /mnt/home/bf996/experiments/densemixer/data_nemotron_code_oracle
    ds_type: parquet
    data_files:
      - /mnt/home/bf996/experiments/densemixer/data_nemotron_code_oracle/data/train-*.parquet
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: role
      content: content
    split_thinking: false
chat_template: chatml
dataset_prepared_path: /mnt/home/bf996/experiments/densemixer/prepared/run1   # SHARED with Run 1 (same tokens + order)
val_set_size: 0.0
dataset_num_proc: 1
dataloader_num_workers: 2
dataloader_prefetch_factor: 2

# === precision — bf16 + flash-attn (IDENTICAL to Run 1) ===
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

# === memory / compute (IDENTICAL to Run 1) ===
# ⚠ Blackwell fix (B-only, functionally inert for A): the STOCK Qwen3MoE experts default to the
# `grouped_mm` kernel -> `torch._grouped_mm`, which is Hopper-only (cc 9.0) and RuntimeErrors on the
# B200 (cc 10.0) at the first step (job 31707). `eager` uses the per-expert F.linear loop (no
# grouped_mm) -> works on Blackwell. This is NOT a control confound: A (dense_mixer:true) replaces the
# whole SparseMoeBlock.forward with the tf-5.x port that accesses expert weights directly and NEVER
# calls self.experts.forward, so `experts_implementation` is never exercised on A's path — the only
# FUNCTIONAL A/B difference remains dense (all-expert STE) vs sparse (top-k). Both do per-expert F.linear.
experts_implementation: eager
deepspeed: /opt/axolotl/deepspeed_configs/zero3_bf16.json
gradient_checkpointing: true
chunked_cross_entropy: true
sequence_len: 16384
sample_packing: true

# === control discipline — IDENTICAL to Run 1 ===
seed: 42
micro_batch_size: 1
gradient_accumulation_steps: 4
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 (IDENTICAL to Run 1) — θ₀ + intermediate + final for the Δθ trajectory ===
logging_steps: 1
save_steps: 10
save_total_limit: 100
output_dir: /mnt/home/bf996/experiments/densemixer/run2_sparse_out   # DISTINCT from Run 1 (not a control var)
special_tokens: {}

```

</details><br>

# mnt/home/bf996/experiments/densemixer/run2_sparse_out

This model was trained from scratch on the None dataset.

## Model description

More information needed

## Intended uses & limitations

More information needed

## Training and evaluation data

More information needed

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 1
- eval_batch_size: 1
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- gradient_accumulation_steps: 4
- total_train_batch_size: 32
- total_eval_batch_size: 8
- optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 9
- training_steps: 96

### Training results



### Framework versions

- Transformers 5.12.1
- Pytorch 2.8.0a0+5228986c39.nv25.06
- Datasets 4.8.5
- Tokenizers 0.23.0-rc0