[inureyes]

Summary

PR #1389 already wired internlm2::Attention to the shared DynamicNtkRope schedule (issue #1324), so the rope_scaling and rope_traditional plumbing itself was already correct on main. This PR closes the rest of #1320: family-level tests proving that plumbing reaches the helper, a debug log of the rescaled base, and real-checkpoint validation.

What changed

• src/models/internlm2_tests.rs: tests that ModelArgs::rope(), not just the helper, resolves a linear block to an inverse-factor scale, carries rope_traditional through in both states, reproduces the issue's dynamic table (base_for(40000), base_for(65536)), accepts the rope_type spelling, and rejects a factor-less dynamic block with an error naming the family.
• src/models/dynamic_ntk_rope.rs: a traditional() accessor, and a debug-level log of base_eff the first time a dynamic schedule crosses max_position_embeddings. The dedup is keyed on the schedule's own parameters rather than a process-wide flag, following rope_utils::report_unusable_rope_scaling_once, so a second checkpoint in the same process (--models-dir routing, pipeline stages, TP ranks) still gets its line. apply asks tracing::enabled! before it takes the dedup lock, so a run with no debug subscriber never touches that shared lock; the rotary arithmetic is unchanged either way.
• Comments and docs/supported-models.md named a checkpoint directory that does not exist (models/internlm2-7b-4bit); corrected to models/internlm2_5-7b-chat-4bit, and the entry now records the long-context evidence rather than crediting the short run with it, at the figures the closed form actually gives.

Validation

--lib models::internlm2 (8 passed) and --lib models::dynamic_ntk_rope (17 passed) under --profile test-fast --features metal,accelerate; cargo clippy --lib --tests and cargo fmt --all --check clean, which is the scope ci.yml runs. The --workspace gate in issue #1320's acceptance criteria runs in nightly-verify.yml, not on this PR.

Real checkpoint models/internlm2_5-7b-chat-4bit, greedy. Against a binary built from this branch's parent commit, an 89-token prompt at -n 64 and a 34021-token prompt at -n 32 both produce byte-identical generated text (MD5 5a6d6133 and 69a12bff). That pair is a control rather than a parity result: the parent already carries #1389's wiring, so the two arms differ only by the debug diagnostic this PR adds, and byte-identity is what it should show.

The long-context run was repeated against the final binary, after the tracing::enabled! gate was added, so the diagnostic is confirmed to still reach a RUST_LOG=debug run rather than only having done so before that commit. A 34021-token prompt emits exactly one line, seq_len=34021 base_eff=1077737.0, and the generation answers the question the prompt ends with in fluent on-topic text (Rayleigh scattering). 1077737.0 is the issue's closed form at that length: f64 gives 1077736.99 there, so the f32 the code computes agrees to every digit the line prints, and a 34022-token prompt gives base_eff=1077799.125, which the same f32 evaluation reproduces bit for bit. The base moves about 62 per token at that length, which is why the entry in docs/supported-models.md now quotes the value with its length; its previous parenthetical read the closed form at 34020 as an f64-versus-f32 gap at 34021. The same form gives the issue's table at 40000 (1449795.74) and 65536 (3052773.67).

Closes #1320

[inureyes]

PR Finalization Check

Reviewed test coverage, documentation, and lint/format on fix/issue-1320-internlm2-rope-scaling at 1aba962d. No gaps found, no changes made.

Tests

--lib models::internlm2 (8 passed) and --lib models::dynamic_ntk_rope (17 passed) under --profile test-fast --features metal,accelerate both pass locally, matching the counts in the PR description. Coverage already spans the linear/dynamic/absent rope_scaling cases, both type and rope_type spellings, rope_traditional true and the serde-default false, the factor-less dynamic load error naming the family, and a dedicated dedup test for the rescale log (first-crossing fires, repeat and per-layer-rebuilt schedules stay silent, a different base or factor still logs, and both the boundary and non-dynamic modes never log). No real gap to add tests for.

Documentation

docs/supported-models.md already carries the corrected checkpoint path and states plainly what the long-context run does and does not establish. Checked docs/llama-server-compat.md and docs/model-catalog.tsv for stale references to the old internlm2-7b-4bit name; the TSV already tracks it as a documented alias of internlm2_5-7b-chat-4bit, and the benchmark result logs under docs/benchmark_results/ are historical run records that correctly use whatever name was live at the time, so neither needed a change. No docs/ko tree exists in this repository (the mkdocs manual under docs/en and docs/ko is a separate documentation tree per the Makefile comments), so there is no Korean file to update.

Lint / format

cargo fmt --all --check and cargo clippy --lib --tests --features metal,accelerate -- -D warnings both pass clean. cargo check --lib --tests also clean.

All PR CI checks (cargo-clippy, cargo-fmt, cargo-deny, OpenXLA feature compile, crate versions, cross-repo refs, kernel dtype keys, llama-compat manifest) are green. Ready for merge from this angle.