[quant][fx] store dtype, axis as literals in the graph

[supriyar]

Stack from ghstack:

• [quant][fx] Add pass in convert to fold quant-dequant sequence #54860 [quant][fx] Add pass in convert to fold quant-dequant sequence
• [quant][fx][pyper] Get first linear use of quantize_per_tensor for FQN #54859 [quant][fx][pyper] Get first linear use of quantize_per_tensor for FQN
• [quant][fx] store dtype, axis as literals in the graph #54624 [quant][fx] store dtype, axis as literals in the graph

Summary:
previously we were creating setattr nodes for dtype and axis.
The FX convention is that primitive types are embedded as literals in args/kwargs.

With this change we won't see getattr nodes in the graph anymore for dtype/axis

Test Plan:
python test/test_quantization.py TestQuantizeFx

Reviewers:

Subscribers:

Tasks:

Tags:

Differential Revision: D27306898

[facebook-github-bot]

💊 CI failures summary and remediations

As of commit 574efc3 (more details on the Dr. CI page):

---

💚 💚 Looks good so far! There are no failures yet. 💚 💚

---

This comment was automatically generated by Dr. CI (expand for details).

Follow this link to opt-out of these comments for your Pull Requests.

Please report bugs/suggestions to the (internal) Dr. CI Users group.

[jerryzh168]

do we need "dtype" to appear in state_dict?

[supriyar]

do we need "dtype" to appear in state_dict?

I don't think so. I think this change existed even before we added support for scale/zero_point to appear in the state_dict. Since we never call register_buffer for dtype, it was never part of state_dict.

[jerryzh168]

do we need "dtype" to appear in state_dict?

I don't think so. I think this change existed even before we added support for scale/zero_point to appear in the state_dict. Since we never call register_buffer for dtype, it was never part of state_dict.

so only scale/zero_point need to be loaded from state_dict, dtype is fixed?

[supriyar]

do we need "dtype" to appear in state_dict?

I don't think so. I think this change existed even before we added support for scale/zero_point to appear in the state_dict. Since we never call register_buffer for dtype, it was never part of state_dict.

so only scale/zero_point need to be loaded from state_dict, dtype is fixed?

Yes, the dtype information is retrieved from the qconfig, so it doesn't need to be stored in the state_dict. As long as we use the same config at load time, it should match.

[jerryzh168]

do we need "dtype" to appear in state_dict?

I don't think so. I think this change existed even before we added support for scale/zero_point to appear in the state_dict. Since we never call register_buffer for dtype, it was never part of state_dict.

so only scale/zero_point need to be loaded from state_dict, dtype is fixed?

Yes, the dtype information is retrieved from the qconfig, so it doesn't need to be stored in the state_dict. As long as we use the same config at load time, it should match.

I see, sounds good.

[codecov]

Codecov Report

Merging #54624 (574efc3) into gh/supriyar/220/base (12a61a1) will decrease coverage by 0.39%.
The diff coverage is 100.00%.

@@                   Coverage Diff                    @@
##           gh/supriyar/220/base   #54624      +/-   ##
========================================================
- Coverage                 77.47%   77.08%   -0.40%     
========================================================
  Files                      1893     1893              
  Lines                    185921   185918       -3     
========================================================
- Hits                     144043   143307     -736     
- Misses                    41878    42611     +733     

[facebook-github-bot]

This pull request has been merged in a0a7a2d.