From 0e68bb995b98544f3cb021c3770699052c286fcf Mon Sep 17 00:00:00 2001 From: saitejagirada Date: Sat, 12 Sep 2026 09:04:34 +0530 Subject: [PATCH] test: add reason strings to bare #[ignore] attributes Fill every bare `#[ignore]` and `#[ignore] // comment` attribute under src/ and tests/ with a reason string in the `#[ignore = "..."]` form, matching the convention used by the other ~350 ignore attributes in the tree. In every case the reason text was already adjacent in the file (a module doc comment, a `MODEL_DIR` const, an `eprintln!` skip message, or a trailing comment). Eight files, 23 sites: - tests/deepseek_v4_real_model.rs (5) - tests/qwen38_mtp_chain_parity.rs (6) - tests/got_ocr_prompt_parity.rs (3) - tests/qwen3_omni_moe_parity.rs (1) - tests/ernie4_5_moe_vl_parity.rs (1) - src/models/diffusion_gemma/tests.rs (2) - src/downloader/tests.rs (1) - src/tokenizer/tiktoken_tests.rs (4) Closes #1698 --- src/downloader/tests.rs | 2 +- src/models/diffusion_gemma/tests.rs | 4 ++-- src/tokenizer/tiktoken_tests.rs | 8 ++++---- tests/deepseek_v4_real_model.rs | 10 +++++----- tests/ernie4_5_moe_vl_parity.rs | 2 +- tests/got_ocr_prompt_parity.rs | 6 +++--- tests/qwen38_mtp_chain_parity.rs | 12 ++++++------ tests/qwen3_omni_moe_parity.rs | 2 +- 8 files changed, 23 insertions(+), 23 deletions(-) diff --git a/src/downloader/tests.rs b/src/downloader/tests.rs index 1eda9151a..fe18aa9e1 100644 --- a/src/downloader/tests.rs +++ b/src/downloader/tests.rs @@ -544,7 +544,7 @@ async fn stream_file_cleans_up_tempfile_on_error() { // Integration test that hits the real Hugging Face Hub. Marked `#[ignore]` // per the issue acceptance criteria so CI does not depend on network access. #[test] -#[ignore] +#[ignore = "hits the real Hugging Face Hub; CI must not depend on network access"] fn live_download_smoke_test() { // Use a tiny repo so the test stays cheap when explicitly requested with // `cargo test -- --ignored`. diff --git a/src/models/diffusion_gemma/tests.rs b/src/models/diffusion_gemma/tests.rs index 92011c707..cb770b4fe 100644 --- a/src/models/diffusion_gemma/tests.rs +++ b/src/models/diffusion_gemma/tests.rs @@ -445,7 +445,7 @@ fn dense_windowed_mask_is_correct_for_multi_token_offset_forward() { /// /// `cargo test --release --lib models::diffusion_gemma::tests::real_model_forward_determinism -- --ignored --nocapture` #[test] -#[ignore] +#[ignore = "real-model regression test; needs the DiffusionGemma checkpoint on disk"] fn real_model_forward_determinism() { let dir = std::path::Path::new(env!("CARGO_MANIFEST_DIR")) .join("models/diffusiongemma-26B-A4B-it-4bit"); @@ -800,7 +800,7 @@ fn sanitize_drops_clip_calibration_when_unclipped() { /// /// `cargo test --release --lib models::diffusion_gemma::tests::real_model_loads_vision_front_end -- --ignored --nocapture` #[test] -#[ignore] +#[ignore = "real-model regression test; needs the DiffusionGemma checkpoint on disk"] fn real_model_loads_vision_front_end() { let dir = std::path::Path::new(env!("CARGO_MANIFEST_DIR")) .join("models/diffusiongemma-26B-A4B-it-4bit"); diff --git a/src/tokenizer/tiktoken_tests.rs b/src/tokenizer/tiktoken_tests.rs index 2a4940b5b..cb9fa7269 100644 --- a/src/tokenizer/tiktoken_tests.rs +++ b/src/tokenizer/tiktoken_tests.rs @@ -117,7 +117,7 @@ fn qwen_image_tags_encode_as_single_specials() { } #[test] -#[ignore] // Requires model files +#[ignore = "requires model files on disk"] fn test_load_tiktoken() { let tokenizer = TiktokenTokenizer::from_file(&tiktoken_file(), &test_model_path()); assert!( @@ -132,7 +132,7 @@ fn test_load_tiktoken() { } #[test] -#[ignore] // Requires model files +#[ignore = "requires model files on disk"] fn test_encode_decode_roundtrip() { let t = TiktokenTokenizer::from_file(&tiktoken_file(), &test_model_path()).unwrap(); let text = "Hello, world!"; @@ -143,7 +143,7 @@ fn test_encode_decode_roundtrip() { } #[test] -#[ignore] // Requires model files +#[ignore = "requires model files on disk"] fn test_encode_chinese() { let t = TiktokenTokenizer::from_file(&tiktoken_file(), &test_model_path()).unwrap(); let text = "你好，世界"; @@ -154,7 +154,7 @@ fn test_encode_chinese() { } #[test] -#[ignore] // Requires model files +#[ignore = "requires model files on disk"] fn test_special_tokens() { let t = TiktokenTokenizer::from_file(&tiktoken_file(), &test_model_path()).unwrap(); let text = "<|eos|>"; diff --git a/tests/deepseek_v4_real_model.rs b/tests/deepseek_v4_real_model.rs index 978b83a3c..ca0854634 100644 --- a/tests/deepseek_v4_real_model.rs +++ b/tests/deepseek_v4_real_model.rs @@ -52,7 +52,7 @@ use mlxcel::{CxxGenerator, LanguageModel, SamplingConfig, initialize_runtime, lo const MODEL_DIR: &str = "deepseek-v4-flash-4bit"; #[test] -#[ignore] +#[ignore = "real-model heavy (~151 GB checkpoint on disk; needs a high-memory Apple Silicon host)"] fn deepseek_v4_real_model_loads_and_generates_coherently() { let model_dir = repo_model_dir(MODEL_DIR); if !model_dir.join("config.json").exists() { @@ -108,7 +108,7 @@ fn deepseek_v4_real_model_loads_and_generates_coherently() { } #[test] -#[ignore] +#[ignore = "real-model heavy (~151 GB checkpoint on disk; needs a high-memory Apple Silicon host)"] fn deepseek_v4_real_model_decode_crosses_pooling_windows() { // A longer decode than one compress window (ratio 4) plus one full // simple window boundary region, so decode-mode pooling emission and the @@ -142,7 +142,7 @@ fn deepseek_v4_real_model_decode_crosses_pooling_windows() { } #[test] -#[ignore] +#[ignore = "real-model heavy (~151 GB checkpoint on disk; needs a high-memory Apple Silicon host)"] fn deepseek_v4_real_model_long_context_hits_sparse_and_compressed_paths() { // The sparse split-softmax and the batched/decode HiSA selection only // run once a ratio-4 layer's pooled count exceeds index_topk (512), i.e. @@ -233,7 +233,7 @@ fn deepseek_v4_real_model_long_context_hits_sparse_and_compressed_paths() { /// world knowledge. Making the answer purely synthetic would test retrieval /// harder and be far less stable on a base checkpoint at this length. #[test] -#[ignore] +#[ignore = "real-model heavy (~151 GB checkpoint on disk; needs a high-memory Apple Silicon host)"] fn deepseek_v4_real_model_long_context_engages_hisa_hierarchy() { let model_dir = repo_model_dir(MODEL_DIR); if !model_dir.join("config.json").exists() { @@ -359,7 +359,7 @@ fn deepseek_v4_real_model_long_context_engages_hisa_hierarchy() { /// crash or a degenerate output in any band shows up. /// #[test] -#[ignore] +#[ignore = "real-model heavy (~151 GB checkpoint on disk; needs a high-memory Apple Silicon host)"] fn deepseek_v4_real_model_hisa_decode_cost_scaling() { let model_dir = repo_model_dir(MODEL_DIR); if !model_dir.join("config.json").exists() { diff --git a/tests/ernie4_5_moe_vl_parity.rs b/tests/ernie4_5_moe_vl_parity.rs index 13e411fc5..8afd36178 100644 --- a/tests/ernie4_5_moe_vl_parity.rs +++ b/tests/ernie4_5_moe_vl_parity.rs @@ -129,7 +129,7 @@ fn text_only_forward_produces_finite_logits() { /// special-token-only degeneration is visible. Ignored by default; run with /// `--ignored --nocapture` when debugging. #[test] -#[ignore] +#[ignore = "diagnostic; needs a real ERNIE-4.5 MoE VL checkpoint on disk"] fn debug_image_greedy_ids() { let Some(dir) = model_dir() else { return }; let (model, tokenizer) = mlxcel::load_model(&dir).expect("load ERNIE-4.5-VL"); diff --git a/tests/got_ocr_prompt_parity.rs b/tests/got_ocr_prompt_parity.rs index 344441191..a5e0395a3 100644 --- a/tests/got_ocr_prompt_parity.rs +++ b/tests/got_ocr_prompt_parity.rs @@ -78,7 +78,7 @@ fn resolve_checkpoint() -> Option { } #[test] -#[ignore] +#[ignore = "needs a GOT-OCR 2.0 checkpoint on disk"] fn got_prompt_ids_match_the_reference_tokenizer() { let Some(model_dir) = resolve_checkpoint() else { eprintln!( @@ -137,7 +137,7 @@ fn got_prompt_ids_match_the_reference_tokenizer() { } #[test] -#[ignore] +#[ignore = "needs a GOT-OCR 2.0 checkpoint on disk"] fn got_format_mode_prompt_ids_match_the_reference_tokenizer() { let Some(model_dir) = resolve_checkpoint() else { eprintln!("Skipping: no GOT-OCR 2.0 checkpoint found."); @@ -164,7 +164,7 @@ fn got_format_mode_prompt_ids_match_the_reference_tokenizer() { /// id. The render is the builtin GOT template's output, which is what /// `ChatTemplateProcessor` hands the runtime for a chat request. #[test] -#[ignore] +#[ignore = "needs a GOT-OCR 2.0 checkpoint on disk"] fn server_render_and_cli_instruction_tokenize_identically() { let Some(model_dir) = resolve_checkpoint() else { eprintln!("Skipping: no GOT-OCR 2.0 checkpoint found."); diff --git a/tests/qwen38_mtp_chain_parity.rs b/tests/qwen38_mtp_chain_parity.rs index be1f50c6e..ff90e8751 100644 --- a/tests/qwen38_mtp_chain_parity.rs +++ b/tests/qwen38_mtp_chain_parity.rs @@ -124,7 +124,7 @@ fn forward_tokens(model: &Qwen35Model, tokens: &[i32], caches: &mut [Qwen3NextCa } #[test] -#[ignore] +#[ignore = "real-model heavy"] fn block_verify_chain_matches_single_token_chain() { let target_dir = CANDIDATE_TARGETS .iter() @@ -199,7 +199,7 @@ fn block_verify_chain_matches_single_token_chain() { /// every GPU is that a passing probe implies a passing chain, because the /// gate turns MTP on off the back of it. #[test] -#[ignore] +#[ignore = "real-model heavy"] fn the_exactness_probe_never_passes_where_the_block_chain_diverges() { let target_dir = CANDIDATE_TARGETS .iter() @@ -444,7 +444,7 @@ fn assert_parity_and_standard_arms_are_distinguishable() { /// cargo test --test qwen38_mtp_chain_parity --release --features metal,accelerate -- --ignored --nocapture parity_kernel_cost /// ``` #[test] -#[ignore] +#[ignore = "real-model heavy"] fn parity_kernel_cost_vs_standard() { if !mlxcel_core::metal_is_available() { eprintln!("skipping: chain-parity kernel is Metal-only"); @@ -520,7 +520,7 @@ fn parity_kernel_cost_vs_standard() { /// cargo test --test qwen38_mtp_chain_parity --release --features metal,accelerate -- --ignored --nocapture verify_forward_cost /// ``` #[test] -#[ignore] +#[ignore = "real-model heavy"] fn verify_forward_cost_scaling() { let Some(target_dir) = CANDIDATE_TARGETS .iter() @@ -609,7 +609,7 @@ fn greedy() -> SamplingConfig { /// runs. Any divergence from the classic chain is therefore in the adapter's /// prefill or the multi-token verify forward itself. #[test] -#[ignore] +#[ignore = "real-model heavy; needs a Qwen 3.5-family checkpoint on disk"] fn adapter_chain_with_perfect_drafts_matches_classic_chain() { let Some(target_dir) = CANDIDATE_TARGETS .iter() @@ -675,7 +675,7 @@ fn adapter_chain_with_perfect_drafts_matches_classic_chain() { /// conditions on the accepted prefix, so its argmax must still follow the /// classic chain; any divergence isolates the rollback path. #[test] -#[ignore] +#[ignore = "real-model heavy; needs a Qwen 3.5-family checkpoint on disk"] fn adapter_chain_with_rejected_drafts_matches_classic_chain() { let Some(target_dir) = CANDIDATE_TARGETS .iter() diff --git a/tests/qwen3_omni_moe_parity.rs b/tests/qwen3_omni_moe_parity.rs index 45a759251..779d4bebc 100644 --- a/tests/qwen3_omni_moe_parity.rs +++ b/tests/qwen3_omni_moe_parity.rs @@ -63,7 +63,7 @@ fn audio_out_len_formula() { /// Diagnostic: greedy-decode a few steps after a TEMPLATED image prefill and /// dump raw ids. Ignored by default. #[test] -#[ignore] +#[ignore = "diagnostic; needs a real Qwen3-Omni checkpoint on disk"] fn debug_templated_image_greedy_ids() { let Some(dir) = model_dir() else { return }; let (model, tokenizer) = mlxcel::load_model(&dir).expect("load qwen3-omni");