fix(rope): tolerate missing original_max_position_embeddings

[lizamd]

Motivation

get_rope() raises KeyError and every engine dies at init when a model's
rope_scaling lacks original_max_position_embeddings:

File "aiter/rotary_embedding.py", line 1872, in get_rope
    original_max_position = rope_scaling["original_max_position_embeddings"]
KeyError: 'original_max_position_embeddings'

Root cause is a transformers v4 -> v5 change. On v4, unscaled models had
config.rope_scaling = None. On v5 it aliases rope_parameters and is always
populated, e.g. for GLM-4.7-Flash:

{'rope_theta': 1000000, 'partial_rotary_factor': 1.0, 'rope_type': 'default'}

sglang's DeepseekV2AttentionMLA treats a truthy rope_scaling as "scaling is
configured" and sets rope_type = "deepseek_yarn", so an unscaled MLA model
reaches the yarn branch without the keys that branch requires.

sglang already handles this in
sglang/srt/layers/rotary_embedding/factory.py::_get_rope_param (warn + default).
aiter's vendored get_rope never received the equivalent, so on ROCm the same
config is a hard failure rather than a warning.

Technical Details

Adds _get_rope_param(rope_scaling, key, default, scaling_type), mirroring
sglang's helper: return the key if present, otherwise log a warning and fall
back to max_position. Applied at all four call sites that read
original_max_position_embeddings (llama3, yarn, deepseek_yarn,
su/longrope) for parity with sglang.

The warning matters: silently defaulting would hide a genuinely mis-specified
config for a model that really does use scaling.

Correctness for the unscaled case: with factor absent (defaults to 1.0) and
original_max_position == max_position, YaRN degenerates to standard RoPE — no
interpolation, and the MLA attention scaling is unchanged. Verified:

compute_mla_mscale_scaling({'rope_type':'deepseek_yarn', ...}, 0.1234) -> 0.1234

Test Plan

• GLM-4.7-Flash (MLA, rope_type='default') on 8x MI355X, ROCm 7.2,
  transformers 5.12.1, sglang 0.5.17, SGLANG_USE_AITER=1
• Start all 8 sglang engines, then run an end-to-end GRPO training loop
  (rollout -> backward -> weight sync)

Test Result

Before: all 8 engines abort at init with the KeyError above.

After: engines initialize (max_total_num_tokens=2870263, context_len=202752)
and the training loop runs end to end, with the expected warning emitted once
per engine. All aiter kernels remain enabled — SGLANG_USE_AITER=1 throughout;
no fallback to non-aiter paths was needed.

Submission Checklist

• Change is limited to the missing-key fallback; no kernel or numerics
  changes for configs that already carry the key
• Verified on MI355X (gfx950)
• Mirrors existing sglang behavior rather than introducing new semantics

[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:gfx1250-ffm-triton	Run the five-shard gfx1250 FFM Triton test suite
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 4742 --add-label <label>