[zhangzhenghao]

Summary

• Add storage-only IEEE binary16 tensors with explicit DType, generic NewTensor[T float32 | uint16], raw Uint16s() / Floats() backing-storage accessors, and independent graph-free ToFloat16 / ToFloat32 conversions.
• Preserve dtype and raw bits through slicing, indexed slicing, cloning, reads, and formatting.
• Serialize both dtypes through shared protobuf bytes data = 3, with element width selected by dtype. Tensor/protobuf payloads share backing storage through unsafe byte views; no dual writes or fallback decoder.
• Keep FP32 operator dispatch, automatic differentiation, optimizers, and initialization unchanged. FP16 tensors must be explicitly converted to FP32 for computation in this PR. No new dtype preflight guards are introduced.
• FP16 operator/autodiff support, SGD/Adam updates, and NormalInit are split into dependent computation PR support FP16 tensor computation through FP32 kernels #1423. That PR targets pr-1398-fp16-storage, a mirror of this PR's head, so its diff contains only computation changes. Retarget it to master after this PR merges.

Verification

Passed locally with system Go 1.27.3 (no scratch modfile or dependency changes):

• go test ./common/nn ./protocol -count=1
• go test -race ./common/nn ./protocol -count=1
• go test -tags=noasm ./common/nn ./protocol -count=1
• go test -gcflags=all=-d=checkptr=2 ./common/nn ./protocol -count=1
• go vet -composites=false ./common/nn ./protocol
• git diff --check

The split retains the original storage commits and subsequent accessor/conversion/serialization changes, while removing the computation commit entirely from this branch's history. The dependent computation tree is byte-for-byte identical to the previous combined PR head. Independent review found no split-induced omissions or blocking issues. GitHub Actions run 40984389040 passed all unit, integration, compatibility, playground, and lint jobs on this head. The separate Codecov project check failed: project coverage 78.17% (-0.93% versus base), patch coverage 91.38% with 5 missing/partial changed lines in tensor.go. This is a coverage check failure, not a test failure; no unrelated coverage changes were bundled into the split.

[codecov]

Codecov Report

❌ Patch coverage is 91.37931% with 5 lines in your changes missing coverage. Please review.
✅ Project coverage is 78.17%. Comparing base (9424736) to head (a36f4b3).
⚠️ Report is 18 commits behind head on master.

Files with missing lines	Patch %	Lines
common/nn/tensor.go	91.07%	4 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1398      +/-   ##
==========================================
- Coverage   79.09%   78.17%   -0.93%     
==========================================
  Files          92       90       -2     
  Lines       16459    16468       +9     
==========================================
- Hits        13019    12874     -145     
- Misses       3440     3593     +153     
- Partials        0        1       +1     

☔ View full report in Codecov by Harness.
📢 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.