[Bug] GPT-OSS raw converter emits incorrect expert weight format, causing gibberish output with `--megatron-to-hf-mode raw`

Bug Description

has two bugs when using --megatron-to-hf-mode raw (non-colocate). After the first weight sync, SGLang produces gibberish (reward=0, 100% truncation).

1. fc1: Emits separate gate_proj/up_proj per expert instead of fused interleaved gate_up_proj. SGLang expects fused 3D tensors via .
2. fc2: Missing transpose — Megatron (out, in) vs HF (in, out). Hidden in 20B where hidden == inter == 2880.

uses --megatron-to-hf-mode bridge --colocate, bypassing entirely, so this path was never tested.

Steps to Reproduce

1. Use origin/main (verified at 8efb1166)
2. Run GPT-OSS 20B GRPO with --megatron-to-hf-mode raw (non-colocate, 2-node: training TP=1/PP=2/EP=4, rollout SGLang TP=2/DP-attn=4)
3. Observe rollout after first weight sync

Expected Behavior

Coherent responses, truncated_ratio < 1%, reward > 0

Actual Behavior

response_len/mean: 2048.0   (all hit max length)
truncated_ratio:   1.0       (100%)
reward:            0

Broken code:

# fc1 — separate unfused (wrong)
gate_weight, up_weight = param.chunk(2, dim=0)
return [
    (f"...experts.{expert_idx}.gate_proj.weight", gate_weight),
    (f"...experts.{expert_idx}.up_proj.weight", up_weight),
]
# fc2 — missing transpose
return [(f"...experts.{expert_idx}.down_proj.weight", param)]

Fix:

# fc1 — fused interleaved
gate, up = param.chunk(2, dim=0)
fused = torch.stack([gate, up], dim=1).reshape(-1, param.shape[-1])
fused = fused.transpose(0, 1).contiguous()
return [(f"...experts.{expert_idx}.gate_up_proj.weight", fused)]
# fc2 — transposed
return [(f"...experts.{expert_idx}.down_proj.weight", param.T.contiguous())]

After fix: response_len/mean: 427, truncated_ratio: 0.004, reward: 0.078, grad_norm: 0.377.

The weight sync pipeline also needs to accumulate per-expert 2D tensors into fused 3D [num_experts, ...] before sending to SGLang, since weight_loader_fused() only accepts 3D shapes.

Environment

slime version: origin/main at 8efb1166

• Python version: 3.10
• PyTorch version: 2.6.0
• CUDA version: 12.8
• GPU type and count: 16x H200
• OS: Ubuntu 22.04 (Docker)
• SGLang version: v0.5.9 (from slimerl/sglang:v0.5.9 base image)
• Megatron-LM version: commit 3714d81d

Logs

Additional Context

No response

Pre-submission Checklist

• I have read the CONTRIBUTING.md and understand the collaboration scope.
• I have read the documentation and my issue is not addressed there.
• I have searched for existing issues and this is not a duplicate.
• I have provided a minimal, reproducible example.

Start in slime/backends/megatron_utils/megatron_to_hf/gpt_oss.py and compare the raw conversion path with SGLang's fused GPT-OSS expert mapping. Trace the weight sync pipeline to verify that per-expert tensors are accumulated into fused 3D tensors before sending them to SGLang. Done means raw non-colocate GPT-OSS sync produces the expected fused expert format, transposed fc2 weights, and coherent rollout metrics.