Remove redundant config_name arg from JAX->Pytorch model conversion script

[tahsinkose]

Fixes #781.

Background

convert_jax_model_to_pytorch.py previously required a --config_name argument (e.g. pi05_base, pi05_droid) to look up a TrainConfig and extract a Pi0Config. This was unnecessary because all the information the conversion actually needs is either inferable from the checkpoint contents or irrelevant to the weight conversion entirely.

---

Approach

Each field of Pi0Config consumed by the script was audited:

1. pi05 (bool)

Determines the model architecture: which projection keys exist (time_mlp_in/time_mlp_out vs state_proj/action_time_mlp_in/action_time_mlp_out) and whether expert normalization layers use adaptive Dense layers or standard RMSNorm scale.

Resolution: reliably detectable from checkpoint contents — "time_mlp_in" is present in the projection params if and only if the checkpoint is pi05. No path-name heuristics needed.

2. action_dim

Used in PI0Pytorch.__init__ to size the projection layers:

  self.action_in_proj = nn.Linear(config.action_dim, action_expert_config.width)
  self.action_out_proj = nn.Linear(action_expert_config.width, config.action_dim)                                                                                                                          
  self.state_proj = nn.Linear(config.action_dim, action_expert_config.width)

Must be correct for load_state_dict to succeed.

Resolution: inferred from the checkpoint — action_in_proj/kernel.shape[0] gives action_dim directly.

3. action_horizon

Only referenced in forward/sample methods (embed_suffix, forward, sample_actions, denoise_step) for attention mask
construction and output slicing. Never used in __init__, so it has no effect on layer sizes or the weight conversion. Notably it also varies across checkpoints (pi05_droid: 15, pi05_libero: 10, default: 50), so using a hardcoded default would silently be wrong — but since it does not affect the conversion at all, it is simply not needed here. The config.json written alongside the converted weights is a human-readable reference only — nothing in the codebase reads it back. So omitting action_horizon from it has no functional impact on inference or finetuning; callers must construct PI0Pytorch(config) with the correct config themselves regardless.
#### 4. paligemma_variant / action_expert_variant

Already hardcoded elsewhere in the script — the PaliGemmaConfig inline class for the vision/language side and
openpi.models.gemma.get_config("gemma_300m") for the expert. Not needed by the config.

Changes

• Removed config_name parameter from main and convert_pi0_checkpoint
• Auto-detect pi05 from checkpoint contents: "time_mlp_in" in initial_params["projection_params"]
• Infer action_dim from checkpoint: action_in_proj_kernel.shape[0]