From feb0a3c6762a7ba2768cfde7f9f29ea0fefe1dcd Mon Sep 17 00:00:00 2001 From: Zhenshan Xie Date: Mon, 14 Sep 2026 13:19:24 -0700 Subject: [PATCH] test(qwen3_5): qualify the remaining published Qwen3.5 widths The family claimed every dense Qwen3.5 width already, but only the 9B had a manifest, so the three smaller published widths were supported without being measured. Each one builds and matches the Hugging Face reference token for token, normalised edit distance 0.0000, with no change to the family. Every new case is premerge. Internal CI requires it: a family-owned case that is added without premerge coverage is rejected by verify_premerge_coverage_is_preserved as removing required coverage. An earlier revision of this change marked the three widths premerge false to keep the premerge suite building one Qwen3.5 model rather than four, which is what that check exists to prevent. The widths are listed in excluded_profiles rather than benchmarked, because the release-performance workload and receipt were collected only for the 9B. Signed-off-by: Zhenshan Xie --- apps/benchmark/performance/release.yaml | 9 ++++++++ .../qwen3_5/tests/manifests/qwen35-0.8b.json | 22 +++++++++++++++++++ .../qwen3_5/tests/manifests/qwen35-2b.json | 22 +++++++++++++++++++ .../qwen3_5/tests/manifests/qwen35-4b.json | 22 +++++++++++++++++++ 4 files changed, 75 insertions(+) create mode 100644 families/qwen3_5/tests/manifests/qwen35-0.8b.json create mode 100644 families/qwen3_5/tests/manifests/qwen35-2b.json create mode 100644 families/qwen3_5/tests/manifests/qwen35-4b.json diff --git a/apps/benchmark/performance/release.yaml b/apps/benchmark/performance/release.yaml index 89f6cf63df..fffe7770f1 100644 --- a/apps/benchmark/performance/release.yaml +++ b/apps/benchmark/performance/release.yaml @@ -69,6 +69,15 @@ excluded_profiles: reason: *mobilenetv4_performance_exclusion - model: mobilenetv4-conv-large reason: *mobilenetv4_performance_exclusion + - model: qwen35-0.8b + reason: &qwen35_width_performance_exclusion >- + Functional and Hugging Face reference-parity qualification is present for + every published dense Qwen3.5 width, but the release-performance workload + and receipt were collected only for the 9B. + - model: qwen35-2b + reason: *qwen35_width_performance_exclusion + - model: qwen35-4b + reason: *qwen35_width_performance_exclusion - model: dpn92-mx-in1k reason: >- Functional and timm reference-parity qualification is present for every diff --git a/families/qwen3_5/tests/manifests/qwen35-0.8b.json b/families/qwen3_5/tests/manifests/qwen35-0.8b.json new file mode 100644 index 0000000000..bdd0d07515 --- /dev/null +++ b/families/qwen3_5/tests/manifests/qwen35-0.8b.json @@ -0,0 +1,22 @@ +{ + "name": "qwen35-0.8b", + "hf_id": "Qwen/Qwen3.5-0.8B", + "bundle": "qwen35-0.8b.bundle", + "family": "qwen3_5", + "task": "text_generation", + "trust_remote_code": true, + "precision": "fp16", + "testcases": [ + { + "name": "qwen35-0.8b", + "premerge": true, + "prompt": "What is the capital of France? Answer in one word.", + "max_new_tokens": 10, + "use_chat_template": true, + "enable_thinking": false, + "reference_precision": "fp32" + } + ], + "max_sequence_length": 256, + "tensor_parallel_size": 1 +} diff --git a/families/qwen3_5/tests/manifests/qwen35-2b.json b/families/qwen3_5/tests/manifests/qwen35-2b.json new file mode 100644 index 0000000000..60e2a46b62 --- /dev/null +++ b/families/qwen3_5/tests/manifests/qwen35-2b.json @@ -0,0 +1,22 @@ +{ + "name": "qwen35-2b", + "hf_id": "Qwen/Qwen3.5-2B", + "bundle": "qwen35-2b.bundle", + "family": "qwen3_5", + "task": "text_generation", + "trust_remote_code": true, + "precision": "fp16", + "testcases": [ + { + "name": "qwen35-2b", + "premerge": true, + "prompt": "What is the capital of France? Answer in one word.", + "max_new_tokens": 10, + "use_chat_template": true, + "enable_thinking": false, + "reference_precision": "fp32" + } + ], + "max_sequence_length": 256, + "tensor_parallel_size": 1 +} diff --git a/families/qwen3_5/tests/manifests/qwen35-4b.json b/families/qwen3_5/tests/manifests/qwen35-4b.json new file mode 100644 index 0000000000..75c5f7825d --- /dev/null +++ b/families/qwen3_5/tests/manifests/qwen35-4b.json @@ -0,0 +1,22 @@ +{ + "name": "qwen35-4b", + "hf_id": "Qwen/Qwen3.5-4B", + "bundle": "qwen35-4b.bundle", + "family": "qwen3_5", + "task": "text_generation", + "trust_remote_code": true, + "precision": "fp16", + "testcases": [ + { + "name": "qwen35-4b", + "premerge": true, + "prompt": "What is the capital of France? Answer in one word.", + "max_new_tokens": 10, + "use_chat_template": true, + "enable_thinking": false, + "reference_precision": "fp32" + } + ], + "max_sequence_length": 256, + "tensor_parallel_size": 1 +}