[All] Added better error messages

[ptrendx]

Description

Added better error messages throughout the codebase in order to provide more information when the error gets triggered.

[greptile-apps]

Greptile Summary

This PR systematically replaces bare assert statements with explicit, typed exceptions (ValueError, RuntimeError, TypeError) across 31 files, and enriches the resulting error messages with actual runtime values (shapes, dtypes, layout names, counts, etc.) to make debugging significantly easier.

Key changes:

• C++ macros (TRANSFORMER_ENGINE_TYPE_SWITCH_ALL and friends): "Invalid type." → descriptive errors listing the actual dtype and the full set of valid types; the previously reported gap where Int16 was missing from TRANSFORMER_ENGINE_TYPE_SWITCH_ALL's expected list is now corrected.
• fused_attn.cpp: New to_string overloads added for NVTE_QKV_Layout and NVTE_QKV_Format; layout/format switch defaults now report the actual enum value and enclosing function name.
• Python (JAX / PyTorch): Hundreds of bare asserts converted to raise ValueError/RuntimeError/TypeError; identity comparisons (is False, is not False, is not True) replaced with idiomatic truthiness checks, addressing patterns flagged in previous review threads.
• Bonus fixes: global_amax_col = None and global_amax_row = None initialisations quietly resolve latent NameError paths noted in a TODO comment; the typo "Bias is implemented for FP4 GEMM." (missing "not") is corrected to "Bias is not supported in NVFP4QuantizerRef.qgemm". One dtype enum value in an error message (quantizer.cpp) is shown as a raw integer instead of a readable type name and should be updated.

Confidence Score: 4/5

• Safe to merge with minor improvements for error message clarity and undefined variable handling.
• This PR is safe to merge as it contains only error-message improvements and exception-type changes with no algorithmic modifications. All conversions from assert to raise occur in error-handling paths only. Two issues identified: (1) one dtype value shown as raw enum integer instead of human-readable string (style/UX), and (2) a latent NameError risk in a pow_2_scales code path (logic) where global_amax_row should be initialized alongside global_amax_col. Both are low-severity but worth addressing.
• transformer_engine/pytorch/csrc/quantizer.cpp (dtype enum value) and transformer_engine/pytorch/custom_recipes/quantization_nvfp4.py (uninitialized global_amax_row)

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[User Calls API] --> B{Pre-condition check}
    B -->|Before PR: bare assert| C[AssertionError\nno context]
    B -->|After PR: explicit raise| D{Which error type?}
    D -->|Wrong value / range| E[ValueError\nshows actual value + expected]
    D -->|Wrong type| F[TypeError\nshows actual type + expected]
    D -->|Runtime / state violation| G[RuntimeError\ndescribes state + hint]
    C --> H[User must read source\nto understand failure]
    E --> I[User sees what value\nfailed and why]
    F --> I
    G --> I

Loading

_{Last reviewed commit: 78227d8}

[greptile-apps]

_{26 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{26 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[ptrendx]

/te-ci

[jberchtold-nvidia]

LGTM from the TE/JAX side! Thanks! I have not reviewed the core or PyTorch files

[jberchtold-nvidia]

nit SR: If you'd like, appending "This usually occurs when TransformerEngine was not installed properly and the shared library cannot be loaded. Please see the documentation troubleshooting for assistance" but as is it's already an improvement over no message

[ptrendx]

/te-ci

[greptile-apps]

The global_amax_col = None initialisation was added to fix a latent bug, but global_amax_row should also be initialised. In the pow_2_scales branch (lines 604–612), neither variable is assigned, so any code path that references global_amax_row afterwards raises a NameError. The original TODO comment (line 609–610) mentions both variables. Since both are returned at line 682, both must be defined:

Suggested change

	global_amax_col = None
	global_amax_col = None
	global_amax_row = None