HansBug/qwen3-8b-spear-i375

qwen3-8b spear-cdeact iter 0000375

PyTorch Distributed Checkpoint (DCP) snapshot from an internal RL training run.

Source

• Original path: terminal-rl_qwen3-8b_8gpu_mixed_dapo_harness-camel-agent_explore_spear_lite_int_cosine120_a57_life0.005_postnorm_cdeact0.02_2026-06-05_155052_0609copy/iter_0000375
• Run started: 2026-06-05 15:50:52
• Iteration: 375

Key training config (parsed from run name)

field	value
base model	qwen3-8b
world size	8 GPU
RL algo	mixed-DAPO
harness	camel-agent
exploration	spear_lite_int, cosine120, a57, life=0.005
extras	post-norm, cdeact=0.02

Files

8 sharded .distcp files (~13.35 GiB each) + .metadata + common.pt + metadata.json. Backend: torch_dist (sharded), torch (common). Total ~107 GiB.

__{rank}_{shard}.distcp   for rank in 0..3, shard in 0..1
.metadata
common.pt
metadata.json

Loading

import torch.distributed.checkpoint as dcp
state_dict = {...}  # template with empty tensors of expected shapes
dcp.load(state_dict, checkpoint_id="path/to/iter_0000375")

Or convert to a single torch.save file:

python -m torch.distributed.checkpoint.format_utils \
    dcp_to_torch_save path/to/iter_0000375 consolidated.pt

Companion run: HansBug/qwen3-8b-mt10-i343.