Default missing deepseek_yarn original_max_position_embeddings instead of raising

[JessicaJiang-123]

Motivation

AITER's deepseek_yarn RoPE path requires rope_scaling["original_max_position_embeddings"] and raises KeyError when the key is missing.

This breaks GLM-4.7-Flash loading on ROCm with SGLANG_USE_AITER=1:

python3 -m sglang.launch_server --model-path zai-org/GLM-4.7-Flash --trust-remote-code --tp 1

  File "sglang/srt/models/glm4_moe_lite.py", line 544, in __init__
    self.self_attn = DeepseekV2AttentionMLA(
  File "sglang/srt/models/deepseek_v2.py", line 1717, in __init__
    self.rotary_emb = get_rope_wrapper(
  File "sglang/srt/layers/rotary_embedding/factory.py", line 437, in get_rope_wrapper
    return wrapper(
  File "aiter/rotary_embedding.py", line 1872, in get_rope
    original_max_position = rope_scaling["original_max_position_embeddings"]
KeyError: 'original_max_position_embeddings'

GLM-4.7-Flash does not define YaRN parameters in its config, but SGLang relabels the generated rope_scaling dict as deepseek_yarn. SGLang's own RoPE implementation already handles this case by defaulting original_max_position_embeddings to max_position; AITER does not.

Technical Details

This PR ports SGLang's _get_rope_param() fallback to AITER and uses it for original_max_position_embeddings in the deepseek_yarn branch. SGLang added the fallback in sgl-project/sglang#17784.

Related code:

1. factory.py::_get_rope_param()
2. factory.py::deepseek_yarn

Test Plan

MI355X / gfx950 / ROCm 7.2.

1. Test deepseek_yarn with and without original_max_position_embeddings.
2. Launch GLM-4.7-Flash with SGLang and AITER enabled.
3. Run GLM-4.7-Flash RL training on 4x MI355X.

Test Result

case	before	after
missing original_max_position_embeddings	KeyError	fallback applied
DeepSeek-style config with the key present	works	unchanged
GLM-4.7-Flash serving	fails during model loading	server starts and serves
GLM-4.7-Flash RL training	cannot start	completes

RL validation completed with TP2 / PP2 / EP2, EAGLE speculative decoding, and MTP.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 4653 --add-label <label>

[Copilot]

Pull request overview

This PR makes AITER’s deepseek_yarn RoPE configuration handling more robust by defaulting a missing rope_scaling["original_max_position_embeddings"] to max_position (with a warning) instead of raising a KeyError, aligning behavior with SGLang and unblocking GLM-4.7-Flash loading when SGLang relabels RoPE scaling as deepseek_yarn.

Changes:

• Add _get_rope_param() helper to fetch rope_scaling parameters with a warning-and-default fallback.
• Use the helper in the deepseek_yarn branch to avoid KeyError when original_max_position_embeddings is absent.
• Wire in aiter.logger for consistent logging.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[JessicaJiang-123]

Hi @junhaha666, could you please help review this PR when you get a chance? This is currently blocking the Miles CI for GLM-4.7-Flash on ROCm. Thanks!