refactor: migrate embedding storage to BF16 format

[zhangzhenghao]

Summary

This PR migrates embedding storage from float32 to BF16 (bfloat16) format for memory efficiency.

Changes

• Item/User label embeddings: Convert float32/float64 slices to BF16 (uint16) storage
• CTR model: Update Embedding.Value type from []float32 to []uint16
• Dataset processing: Add support for direct []float32 and []float64 label input
• Tests: Update test cases to use BF16 format for embedded labels

Related PRs

• Implement BF16 euclidean #1245: Implement BF16 euclidean distance
• fix: use truncation instead of rounding in ToBF16 #1244: Fix BF16 conversion (truncation instead of rounding)
• feat(master): store item embeddings as BF16 in LoadDataFromDatabase #1238: Store item embeddings as BF16 in LoadDataFromDatabase

Benefits

• ~50% memory reduction for embedding storage
• Maintains sufficient precision for ML workloads
• Consistent BF16 usage across the codebase

[codecov]

Codecov Report

❌ Patch coverage is 69.86301% with 22 lines in your changes missing coverage. Please review.
✅ Project coverage is 73.28%. Comparing base (2e73b7c) to head (a5ddae4).
⚠️ Report is 2 commits behind head on master.

Files with missing lines	Patch %	Lines
common/bfloats/bfloats.go	69.56%	14 Missing ⚠️
model/ctr/data.go	60.00%	4 Missing ⚠️
logics/item_to_item.go	76.92%	2 Missing and 1 partial ⚠️
master/tasks.go	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1246      +/-   ##
==========================================
+ Coverage   73.01%   73.28%   +0.27%     
==========================================
  Files          91       91              
  Lines       16694    16704      +10     
==========================================
+ Hits        12189    12242      +53     
+ Misses       3262     3250      -12     
+ Partials     1243     1212      -31     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull request overview

This PR migrates embedding storage across CTR and item-to-item logic from []float32/[]float64 to BF16 bit-pattern storage ([]uint16) to reduce memory usage and align embedding handling across the codebase.

Changes:

• Update CTR Embedding.Value to []uint16 and propagate through AFM batch prediction code paths.
• Convert dataset label embedding processing to store embeddings as BF16 ([]uint16) and add a BF16 FromAny helper for mixed numeric inputs.
• Update unit tests to assert against BF16-encoded embeddings.

Reviewed changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
model/ctr/model_test.go	Updates CTR model tests to pass BF16 embeddings ([]uint16).
model/ctr/fm_xla.go	Stops converting embeddings to BF16 inside batch prediction (now already BF16).
model/ctr/fm.go	Same as above for the non-XLA AFM path; expects BF16 embeddings.
model/ctr/data_test.go	Updates embedding conversion tests to assert BF16 output.
model/ctr/data.go	Changes Embedding.Value to []uint16 and converts input embedding representations to BF16.
master/tasks.go	Stores item embeddings directly as BF16 in CTR dataset construction.
logics/item_to_item_test.go	Removes toFloat32Slice tests (conversion responsibility moved).
logics/item_to_item.go	Switches embedding item-to-item ANN distance to BF16 Euclidean and uses BF16 vectors.
dataset/dataset_test.go	Updates dataset item label expectations to BF16-encoded embeddings; adds mixed-type input case.
dataset/dataset.go	Converts numeric []any label slices to BF16 during label processing.
common/bfloats/bfloats_test.go	Adds tests for bfloats.FromAny.
common/bfloats/bfloats.go	Adds FromAny and numeric conversion helper for BF16 ingestion.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.