[PyTorch][torch.compile] Make quantizers opaque value objects - #7

Open

pggPL wants to merge 44 commits into

mainfrom

make_qunatizers_opaque

Commits on Jun 29, 2026

[PyTorch] Make tensorless quantizers opaque value objects for torch.compile
pggPL
committed
f3401df
[PyTorch] Drop quantizer value registry; reconstruct via __fx_repr__ globals
pggPL
committed
c4ad54c
[PyTorch] Split dynamo.py into a dynamo/ package
pggPL
committed
a06324b
[PyTorch] Raise in quantizer __fx_repr__ when a process group is stored
pggPL
committed
ea5b396
[PyTorch] Cover NVFP4 in quantizer value-object test
pggPL
committed
aa65e34
Reject a value quantizer that carries an amax reduction group in __eq__/__hash__

pggPL
and
claude
committed
e1b1db6
Recognize value-opaque quantizers via a class flag

pggPL
and
claude
committed
8c33d0e
Address review: narrow opaque-type except, add fullgraph test, fix nvfp4 value key

pggPL
and
claude
committed
945f62d
Restore NVFP4 rht_matrix on value-key rebuild; assert quantize round-trip

pggPL
and
claude
committed
e3c8f43
Enforce process-group rejection in _value_key, not __fx_repr__; add test

pggPL
and
claude
committed
3f68621
Strengthen fullgraph test: quantize/dequantize via a custom op, not passthrough

pggPL
and
claude
committed
32d1768
Clarify comments: rht_matrix_random_sign_mask_t derivation; why the opaque flag

pggPL
and
claude
committed
28bde9e
Reword opaque-flag comment: self-contained, no Linear reference

pggPL
and
claude
committed
2c3c5df
Cover is_opaque_value_type with the import-safety guard too

pggPL
and
claude
committed
826f271
[PyTorch] Expert Parallelism: PyTorch wrapper + autograd ops with symm-mem zero-copy (#3035)
phu0ngng
authored
4cd244e

Commits on Jun 30, 2026

[Common] Update NCCL submodule to have the fix for MAX_SUPPORTED_TOKENS_PER_RANK (#3150)
phu0ngng
authored
90baf02
[PyTorch] Preserve fprop operands for dequantized backward override (#3141)

authored
46bdc85
[PyTorch] Make quantized-tensor __repr__ safe (#3146)

authored
353206d
[Common] EP C API: version config structs and extend `nvte_ep_prepare` with `total_recv_tokens_per_rank` placeholder (#3154)
phu0ngng
authored
3df5e19
Stamp value-opaque flag only after successful registration

pggPL
and
claude
committed
613c545
Drop verbose comments around value-opaque flag stamping

pggPL
and
claude
committed
9db604f
Narrow value process-group check to amax_reduction_group

pggPL
and
claude
committed
3011dfd
Shorten amax_reduction_group check comment

pggPL
and
claude
committed
fe5e5db
Merge remote-tracking branch 'upstream/main' into make_qunatizers_opaque
pggPL
committed
f6b6d78
Drop trivial value-equality boilerplate from quantizer test

pggPL
and
claude
committed
6f66c3e

Commits on Jul 1, 2026

Commits on Jul 2, 2026

docs: document attention backend selection (#3142)
sbhavani
authored
f721112

Commits on Jul 3, 2026

Commits on Jul 6, 2026

Address review comments: qualname registry, import and comment cleanups
pggPL
committed
6fb6a0d
Derive quantizer value fields from class annotations
pggPL
committed
f9c0e18
Drop redundant annotation-exclusion comments
pggPL
committed
dd96956
Shorten _is_value_quantizer comment
pggPL
committed
447a4e1
Fold annotation walk into _value_fields, trim comments
pggPL
committed
02a8fc9
Simplify qualname-registry comment
pggPL
committed
5f55a51
Reword qualname-registry comment for outside readers
pggPL
committed
02afbdc
Fix review findings: pickle compat, subclass opt-in, cached value fields
pggPL
committed
1590931
Validate value-field annotations by resolved type, not annotation text
pggPL
committed
43a4083
Drop the amax_reduction_group fixup from the generic rebuilder
pggPL
committed
10994b7
Merge remote-tracking branch 'upstream/main' into make_qunatizers_opaque
pggPL
committed
f4ddb51
Enable post-RHT amax in the NVFP4 value-object test factory
pggPL
committed
ba520ca