Verify PR #2078 builder design changes already present

[Copilot]

Task: port the design changes (not new models) from onnxruntime-genai#2078 under src/python/py/models/builders/ into modelbuilder/builders/.

After a line-by-line audit of the PR diff against the current tree, every design change is already incorporated. No code modifications are needed.

Items verified as present

• builders/__init__.py — only re-exports Model; model classes are lazy-imported in builder.py.
• base.py
  • Helpers _rope_theta_from_config, _make_search_section, make_neg.
  • make_reshape / make_concat / make_add / make_mul / make_transpose / make_softmax return their output name; make_reshape accepts an inline list shape; make_slice accepts starts=/ends=/axes= kwargs.
  • epsilon = getattr(config, "rms_norm_eps", getattr(config, "layer_norm_eps", 1e-6)).
  • compression_ratio rope rescale lifted from ernie.py into base.
  • make_rope_init honors rope_theta inside rope_scaling, falls back from type to rope_type, threads the truncate flag.
  • make_mscale_yarn(mscale, alpha=1.0) and truncate floor/ceil in make_inv_freq_rescaled_with_ntk.
  • GQA valid set includes ("cpu", FLOAT16).
  • make_attention_input_proj initializes q_path/k_path/v_path; make_repeat_kv shapes use total_sequence_length.
  • cache_dir cleanup guarded by os.path.exists; except Exception as e: print(...) in make_genai_config.
• gemma.py — Gemma3Model records _original_architecture, reads RoPE from rope_parameters (v5+ transformers), and overrides load_weights for Gemma3ForConditionalGeneration.
• gptoss.py — inv_freq = 1.0 / (theta ** ...) correction; gate_up_proj / down_proj transposed before initializer (is_transposed layout); swiglu_limit is None guards; post-MoE cast back to io_dtype.
• nemotron.py — NemotronModel picks up config.norm_eps when present.
• olmo.py — OLMo2Model (post-norm, full-D q/k norm) and OLMo3Model (per-layer sliding window via layer_types).
• smollm.py — saves/restores attention_attrs["rope"]; use_rope_in_attn = has_rope and original_use_rope.
• ernie.py — redundant compression_ratio block removed (handled in base).
• mistral.py / qwen.py — top-level transformers imports trimmed; Mistral3ForConditionalGeneration, Qwen2_5_VLForConditionalGeneration, Qwen3VLForConditionalGeneration, Qwen3_5ForConditionalGeneration are lazy-imported inside load_weights.

New model classes from the PR (HunyuanDenseV1Model, LFM2Model) are intentionally excluded per the issue.

Verification

from modelbuilder.builders.base import Model
import inspect

assert hasattr(Model, "_rope_theta_from_config")
assert hasattr(Model, "_make_search_section")
assert hasattr(Model, "make_neg")
assert {"starts", "ends", "axes"} <= set(inspect.signature(Model.make_slice).parameters)
assert "alpha" in inspect.signature(Model.make_mscale_yarn).parameters

[xadupre]

@copilot there were some design changes requested on the PR linked in the title, can you bring them?

[Copilot]

@copilot there were some design changes requested on the PR linked in the title, can you bring them?

I re-audited the full diff of PR #2078 against modelbuilder/builders/* and every design change in the builders subfolder already appears to be incorporated here (base.py helpers _rope_theta_from_config/_make_search_section/make_neg, return-value plumbing, make_slice kwargs, make_mscale_yarn(alpha=), NTK truncate flag, (cpu, FP16) GQA, total_sequence_length repeat_kv, q/k/v_path init, cache-dir guard, gptoss inv_freq fix + transposed expert weights + swiglu_limit guards + cast-back, OLMo2/OLMo3 with make_genai_config, NemotronH + Nemotron norm_eps, SmolLM3 rope save/restore, Ernie compression_ratio cleanup, Gemma3 _original_architecture + rope_parameters + VLM load_weights, mistral/qwen lazy transformers imports, etc.).

I can't access the PR's review threads (SAML protects the microsoft org from my token), so I may be missing a reviewer-requested change that didn't make it into the final diff. Could you point me at the specific design change you'd like brought over (a thread link, file/line, or short description)?

Warning

Firewall rules blocked me from connecting to one or more addresses (expand for details)

I tried to connect to the following addresses, but was blocked by firewall rules:

• https://api.github.com/repos/microsoft/onnxruntime-genai/pulls/2078/comments
  • Triggering command: /home/REDACTED/work/_temp/ghcca-node/node/bin/node /home/REDACTED/work/_temp/ghcca-node/node/bin/node --enable-source-maps /home/REDACTED/work/_temp/copilot-developer-action-main/dist/index.js (http block)

If you need me to access, download, or install something from one of these locations, you can either:

• Configure Actions setup steps to set up my environment, which run before the firewall is enabled
• Add the appropriate URLs or hosts to the custom allowlist in this repository's Copilot coding agent settings (admins only)