Add `OpFromGraph` wrapper around `alloc_diag` #915

jessegrabowski · 2024-07-10T12:15:44Z

Description

The relatively complex graph generated by pt.diag(x) where x.ndim == 1 is causing problems for rewrites like #860 and #782. We were using a pattern rewriter, but this is proving to be quite brittle. It would be much easier if there were a single Op we could just look for. This PR introduces such an Op, AllocDiag2, as an OpFromGraph wrapper around the existing alloc_diag function.

Related Issue

Closes #
Related to #

Checklist

Checked that the pre-commit linting/style checks pass
Included tests that prove the fix is effective or that the new feature works
Added necessary documentation (docstrings and/or example notebooks)
If you are a pro: each commit corresponds to a relevant logical change

Type of change

…rectly passing

…other rewrites that need it have fired.

jessegrabowski · 2024-07-10T18:07:14Z

I guess my little __props__ thing doesn't work the way I want it to, because these arguments don't get propagated into make_node then the OpFromGraph needs to be rebuilt (I guess that's why the scan tests are failing). I think scan rewrites happen first, so I can't move up the rewrite that eliminates the OpFromGraph. I think the approach here is fundamentally flawed, so need some guidance on how to proceed.

pytensor/tensor/basic.py

ricardoV94 · 2024-07-10T19:35:41Z

We probably want to stop wrapping the gradient in type(self) Since we are using subclass to target specialized forms, we don't want to mistakenly target the gradients. Perhaps the gradient should always use the base OpFromGraph for wrapping? And maybe this wrapping the gradient shouldn't be the default anyway.

>       lop_op = type(self)(
            inputs=inner_inputs + connected_inner_outputs + connected_output_grads,
            outputs=connected_input_grads,
            inline=self.is_inline,
            name=(None if self.name is None else f"{self.name}_LOp"),
            # TODO: We can be eager here and exclude unused inputs in the OFG
            on_unused_input="ignore",
        )
E       TypeError: AllocDiag2.__init__() missing 3 required keyword-only arguments: 'offset', 'axis1', and 'axis2'

jessegrabowski · 2024-07-11T07:28:22Z

Is there a way to just forward all the props there for now? Or do you want me to look into the better solution

ricardoV94 · 2024-07-11T07:40:54Z

Is there a way to just forward all the props there for now? Or do you want me to look into the better solution

I think we should use the BaseClass. Don't want to call AllocDiag to the gradient as well do we?

ricardoV94 · 2024-07-11T07:41:42Z

OpFromGraph wasn't really designed originally to be subclassed so we are doing something new these days, it's not surprising we need to rethink some of the aspects

ricardoV94 · 2024-07-11T07:42:51Z

It also has some sketchy forwarding of "kwargs" already, but those are the compilation kwargs, which is also an odd design (we don't need to tackle that now)

This reverts commit f6f27ec.

ricardoV94

I left some specialized comments, that are only relevant after the big picture question:

Why do we need to inline it specifically for JAX?

pytensor/link/jax/dispatch/tensor_basic.py

codecov · 2024-07-12T06:05:26Z

Codecov Report

Attention: Patch coverage is 97.22222% with 2 lines in your changes missing coverage. Please review.

Project coverage is 81.49%. Comparing base (05d376f) to head (56a3ffe).
Report is 152 commits behind head on main.

Files with missing lines	Patch %	Lines
pytensor/tensor/rewriting/linalg.py	87.50%	1 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #915      +/-   ##
==========================================
+ Coverage   81.04%   81.49%   +0.45%     
==========================================
  Files         170      176       +6     
  Lines       46962    46925      -37     
  Branches    11507    11428      -79     
==========================================
+ Hits        38059    38242     +183     
+ Misses       6694     6500     -194     
+ Partials     2209     2183      -26

Files with missing lines	Coverage Δ
pytensor/compile/builders.py	`88.66% <100.00%> (+0.24%)`	⬆️
pytensor/link/jax/dispatch/basic.py	`80.76% <100.00%> (+4.20%)`	⬆️
pytensor/tensor/basic.py	`91.38% <100.00%> (+2.97%)`	⬆️
pytensor/tensor/elemwise.py	`88.38% <100.00%> (-0.04%)`	⬇️
pytensor/tensor/rewriting/ofg.py	`100.00% <100.00%> (ø)`
pytensor/tensor/rewriting/linalg.py	`90.78% <87.50%> (+1.70%)`	⬆️

... and 79 files with indirect coverage changes

jessegrabowski · 2024-07-12T06:07:57Z

We don't need to only inline it for JAX. I did it only for JAX because we must do it for JAX. Nothing stops us from doing it for all backends, and for all OpFromGraphs. Is there a benefit to doing so?

ricardoV94 · 2024-07-12T06:20:25Z

Why do we need to do it for JAX at all?

jessegrabowski · 2024-07-12T06:24:49Z

Otherwise we need a jax dispatch for AllocDiag specifically

ricardoV94 · 2024-07-12T06:30:51Z

JAX should have a dispatch for OpFromGraph already, does it not?

ricardoV94 · 2024-07-12T06:32:43Z

We do not!!! That's silly, should be as simple as

pytensor/pytensor/link/numba/dispatch/basic.py

Lines 429 to 452 in e2f9cb8

    
           @numba_funcify.register(OpFromGraph) 
        
           def numba_funcify_OpFromGraph(op, node=None, **kwargs): 
        
               _ = kwargs.pop("storage_map", None) 
        
               # Apply inner rewrites 
        
               # TODO: Not sure this is the right place to do this, should we have a rewrite that 
        
               #  explicitly triggers the optimization of the inner graphs of OpFromGraph? 
        
               #  The C-code defers it to the make_thunk phase 
        
               NUMBA.optimizer(op.fgraph) 
        
               fgraph_fn = numba_njit(numba_funcify(op.fgraph, **kwargs)) 
        
               if len(op.fgraph.outputs) == 1: 
        
                   @numba_njit 
        
                   def opfromgraph(*inputs): 
        
                       return fgraph_fn(*inputs)[0] 
        
               else: 
        
                   @numba_njit 
        
                   def opfromgraph(*inputs): 
        
                       return fgraph_fn(*inputs) 
        
               return opfromgraph

jessegrabowski · 2024-07-12T06:33:16Z

Ok I'll implement that instead

jessegrabowski · 2024-07-17T04:54:54Z

Numba segfaulting in the same way I observed in the pad rewrite. I think it's related to the symbolic offset, because this test was passing before. Maybe there's an illegal indexer in the generated code?

pytensor/compile/mode.py

pytensor/tensor/rewriting/linalg.py

pytensor/tensor/rewriting/ofg.py

tests/tensor/rewriting/test_linalg.py

tests/tensor/rewriting/test_ofg.py

ricardoV94 · 2024-07-17T06:03:53Z

Numba segfaulting in the same way I observed in the pad rewrite. I think it's related to the symbolic offset, because this test was passing before. Maybe there's an illegal indexer in the generated code?

Which test is segfaulting?

jessegrabowski · 2024-07-17T06:18:33Z

Looks like case 3 of test_tensor_basic.py::test_ExtractDiag

ricardoV94 · 2024-07-17T06:20:07Z

Looks like case 3 of test_tensor_basic.py::test_ExtractDiag

Can you check how the optimized graph looks like after and before this PR (you have to do this before the linking part starts and segfaults)?

pytensor/tensor/rewriting/linalg.py

pytensor/tensor/rewriting/ofg.py

ricardoV94

Tiny suggestions, looks pretty good.

pytensor/tensor/elemwise.py

…imShuffle`

jessegrabowski · 2024-07-18T05:13:25Z

Follow up on the numba segfault: it appears to happen because fast_run optimizations are omitted from the numba_mode used by compare_numba_and_py. Here's the graph I get when mode="NUMBA":

AdvancedSetSubtensor [id A] 3
 ├─ Alloc [id B] 2
 │  ├─ 0.0 [id C]
 │  ├─ Shape_i{0} [id D] 0
 │  │  └─ <Vector(float64, shape=(?,))> [id E]
 │  └─ Shape_i{0} [id D] 0
 │     └─ ···
 ├─ <Vector(float64, shape=(?,))> [id E]
 ├─ ARange{dtype='int64'} [id F] 1
 │  ├─ 0 [id G]
 │  ├─ Shape_i{0} [id D] 0
 │  │  └─ ···
 │  └─ 1 [id H]
 └─ ARange{dtype='int64'} [id F] 1
    └─ ···

And here's the graph I get when mode = numba_mode:

AllocDiag{axis1=0, axis2=1} [id A]
 ← AdvancedSetSubtensor [id D]
    ├─ Alloc [id E]
    │  ├─ 0.0 [id F]
    │  ├─ Composite{(i1 + abs(i0))} [id G]
    │  │  ├─ *1-<Scalar(int8, shape=())> [id H]
    │  │  └─ Shape_i{0} [id I]
    │  │     └─ *0-<Vector(float64, shape=(?,))> [id J]
    │  └─ Composite{(i1 + abs(i0))} [id G]
    │     └─ ···
    ├─ *0-<Vector(float64, shape=(?,))> [id J]
    ├─ Add [id K]
    │  ├─ ARange{dtype='int64'} [id L]
    │  │  ├─ 0 [id M]
    │  │  ├─ Shape_i{0} [id I]
    │  │  │  └─ ···
    │  │  └─ 1 [id N]
    │  └─ Composite{maximum(0, (-i0))} [id O]
    │     └─ ExpandDims{axis=0} [id P]
    │        └─ *1-<Scalar(int8, shape=())> [id H]
    └─ Add [id Q]
       ├─ ARange{dtype='int64'} [id L]
       │  └─ ···
       └─ Scalarmaximum [id R]
          ├─ [0] [id S]
          └─ ExpandDims{axis=0} [id P]
             └─ ···

Composite{(i1 + abs(i0))} [id G]
 ← add [id T] 'o0'
    ├─ i1 [id U]
    └─ Abs [id V]
       └─ i0 [id W]

Composite{maximum(0, (-i0))} [id O]
 ← maximum [id X] 'o0'
    ├─ 0 [id Y]
    └─ neg [id Z]
       └─ i0 [id BA]

So the OFG isn't written away. Given what I was seeing in the pad PR, my guess is this is something to do with how the inner graphs are being handled when an OFG is jitted.

Or maybe not? When I make the offset arg non-symbolic again, I get this graph (which does not segfault):

AllocDiag{axis1=0, axis2=1} [id A] 0
 └─ <Vector(float64, shape=(?,))> [id B]

Inner graphs:

AllocDiag{axis1=0, axis2=1} [id A]
 ← AdvancedSetSubtensor [id C]
    ├─ Alloc [id D]
    │  ├─ 0.0 [id E]
    │  ├─ Shape_i{0} [id F]
    │  │  └─ *0-<Vector(float64, shape=(?,))> [id G]
    │  └─ Shape_i{0} [id F]
    │     └─ ···
    ├─ *0-<Vector(float64, shape=(?,))> [id G]
    ├─ ARange{dtype='int64'} [id H]
    │  ├─ 0 [id I]
    │  ├─ Shape_i{0} [id F]
    │  │  └─ ···
    │  └─ 1 [id J]
    └─ ARange{dtype='int64'} [id H]
       └─ ···

ricardoV94 · 2024-07-18T05:43:39Z

That maximum and scalar_maximum are a bit suspect. Why is one a ScalarOp and not the other?

Can you dprint with types and destroy/view maps?

jessegrabowski · 2024-07-18T05:49:59Z

without FAST_RUN :

AllocDiag{axis1=0, axis2=1} [id A] <Matrix(float64, shape=(?, ?))> 0
 ├─ <Vector(float64, shape=(?,))> [id B] <Vector(float64, shape=(?,))>
 └─ 0 [id C] <Scalar(int8, shape=())>

Inner graphs:

AllocDiag{axis1=0, axis2=1} [id A]
 ← AdvancedSetSubtensor [id D] <Matrix(float64, shape=(?, ?))> d={0: [0]}
    ├─ Alloc [id E] <Matrix(float64, shape=(?, ?))>
    │  ├─ 0.0 [id F] <Scalar(float64, shape=())>
    │  ├─ Composite{(i1 + abs(i0))} [id G] <Scalar(int64, shape=())> d={0: [1]}
    │  │  ├─ *1-<Scalar(int8, shape=())> [id H] <Scalar(int8, shape=())>
    │  │  └─ Shape_i{0} [id I] <Scalar(int64, shape=())>
    │  │     └─ *0-<Vector(float64, shape=(?,))> [id J] <Vector(float64, shape=(?,))>
    │  └─ Composite{(i1 + abs(i0))} [id G] <Scalar(int64, shape=())> d={0: [1]}
    │     └─ ···
    ├─ *0-<Vector(float64, shape=(?,))> [id J] <Vector(float64, shape=(?,))>
    ├─ Add [id K] <Vector(int64, shape=(?,))>
    │  ├─ ARange{dtype='int64'} [id L] <Vector(int64, shape=(?,))>
    │  │  ├─ 0 [id M] <Scalar(int8, shape=())>
    │  │  ├─ Shape_i{0} [id I] <Scalar(int64, shape=())>
    │  │  │  └─ ···
    │  │  └─ 1 [id N] <Scalar(int8, shape=())>
    │  └─ Composite{maximum(0, (-i0))} [id O] <Vector(int8, shape=(1,))>
    │     └─ ExpandDims{axis=0} [id P] <Vector(int8, shape=(1,))> v={0: [0]}
    │        └─ *1-<Scalar(int8, shape=())> [id H] <Scalar(int8, shape=())>
    └─ Add [id Q] <Vector(int64, shape=(?,))> d={0: [0]}
       ├─ ARange{dtype='int64'} [id L] <Vector(int64, shape=(?,))>
       │  └─ ···
       └─ Scalarmaximum [id R] <Vector(int8, shape=(1,))> d={0: [1]}
          ├─ [0] [id S] <Vector(int8, shape=(1,))>
          └─ ExpandDims{axis=0} [id P] <Vector(int8, shape=(1,))> v={0: [0]}
             └─ ···

Composite{(i1 + abs(i0))} [id G] d={0: [1]}
 ← add [id T] <int64> 'o0'
    ├─ i1 [id U] <int64>
    └─ Abs [id V] <int8>
       └─ i0 [id W] <int8>

Composite{maximum(0, (-i0))} [id O]
 ← maximum [id X] <int8> 'o0'
    ├─ 0 [id Y] <int8>
    └─ neg [id Z] <int8>
       └─ i0 [id BA] <int8>

with FAST_RUN:

AdvancedSetSubtensor [id A] <Matrix(float64, shape=(?, ?))> d={0: [0]} 3
 ├─ Alloc [id B] <Matrix(float64, shape=(?, ?))> 2
 │  ├─ 0.0 [id C] <Scalar(float64, shape=())>
 │  ├─ Shape_i{0} [id D] <Scalar(int64, shape=())> 0
 │  │  └─ <Vector(float64, shape=(?,))> [id E] <Vector(float64, shape=(?,))>
 │  └─ Shape_i{0} [id D] <Scalar(int64, shape=())> 0
 │     └─ ···
 ├─ <Vector(float64, shape=(?,))> [id E] <Vector(float64, shape=(?,))>
 ├─ ARange{dtype='int64'} [id F] <Vector(int64, shape=(?,))> 1
 │  ├─ 0 [id G] <Scalar(int8, shape=())>
 │  ├─ Shape_i{0} [id D] <Scalar(int64, shape=())> 0
 │  │  └─ ···
 │  └─ 1 [id H] <Scalar(int8, shape=())>
 └─ ARange{dtype='int64'} [id F] <Vector(int64, shape=(?,))> 1
    └─ ···

ricardoV94 · 2024-07-18T05:58:11Z

That ScalarMaximum shape=(1,) seems to be trying to write on what's ultimate a shape=() array. Could that be a problem?

Wonder if it's a problem in the C backend. Would need something like Mode(linker=get_mode(”FAST_RUN").linker, optimizer=get_mode(”FAST_COMPILE"). optimizer).

ricardoV94 · 2024-07-18T06:01:02Z

Or can you exclude inplace when optimizing the inner OFG to see if it's related to the problem?

jessegrabowski · 2024-07-18T06:57:50Z

I'm curious why the two maximum(0, offset) and maximum(0, -offset) are being treated so differently

ricardoV94 · 2024-07-18T06:59:40Z

I'm curious why the two maximum(0, offset) and maximum(0, -offset) are being treated so differently

I'm not sure they are, maybe it's one of those Elemwise with a weird custom name. The negated one gets fused with the negation

ricardoV94 · 2024-07-18T07:00:40Z

You can also check what happens with a switch instead of maximum

jessegrabowski · 2024-07-18T07:45:37Z

No problem in the C backend using Mode(linker=get_mode("FAST_RUN").linker, optimizer=get_mode("FAST_COMPILE").optimizer).

Doing numba_mode.excluding('inplace') still segfaults.

Switch also segfaults, as does clip.

jessegrabowski · 2024-07-18T07:46:58Z

Honestly my inclination now is to just switch offset back to a constant, and add a TODO to make it symbolic. I want to get this merged to unblock @tanish1729 's GSOC work.

Edit: Or just change the test to use mode='NUMBA", which works fine. It's only when you use numba_mode = Mode( NumbaLinker(), opts.including("numba", "local_useless_unbatched_blockwise") ) that there's a problem. The default numba mode also includes "fast_run"

ricardoV94 · 2024-07-18T08:17:46Z

Are you excluding inplace "inside" the OFG? That's what I meant for testing inplace

The mode example I gave was to see if the same problem would happen in the C backend. Can't just test with FAST_COMPILE because that skips C

We can make it static no need to fix in this PR. It will come back to us sooner or later.

* Add `OpFromGraph` wrapper around `alloc_diag` * Remove depreciated `AllocDiag` `Op`, rename `AllocDiag2 -> AllocDiag` * Set `inline = False` * Add rewrite to inline all `OpFromGraph` `Op`s * Add `is_zero_offset` helper to `Eye` * Add `is_left_expand_dims` and `is_right_expand_dims` attributes to `DimShuffle` * Seed `test_local_lift_through_linalg` test

Add OpFromGraph wrapper around alloc_diag

6857bea

jessegrabowski added Op implementation OpFromGraph labels Jul 10, 2024

jessegrabowski requested a review from ricardoV94 July 10, 2024 12:15

jessegrabowski added 4 commits July 10, 2024 20:58

Refactor rewrite_det_diag_to_prod_diag to use AllocDiag2

5604d9a

Save arguments passed to alloc_diag as properties in AllocDiag2

b0abe17

Fix bug in rewrite_det_diag_to_prod_diag where batch case was incor…

dbfe92c

…rectly passing

Remove AllocDiag2 from graphs in the specialization phase, after …

afe2a65

…other rewrites that need it have fired.

jessegrabowski force-pushed the alloc-diag-ofg branch from 02fc20c to afe2a65 Compare July 10, 2024 17:53

Remove debugging code, formatting

e810df0

ricardoV94 reviewed Jul 10, 2024

View reviewed changes

pytensor/tensor/basic.py Outdated Show resolved Hide resolved

jessegrabowski added 5 commits July 12, 2024 12:46

Remove depreciated AllocDiag Op, rename AllocDiag2 -> AllocDiag

a5355b6

Correctly register eagerly_inline_alloc_diag as a JAX-only rewrite

6e37d26

Use self (not type(self)) in OpFromGraph.make_node

f6f27ec

Use base class OpFromGraph when constructing OpFromGraph gradients

9486c64

Revert "Use self (not type(self)) in OpFromGraph.make_node"

abcde3f

This reverts commit f6f27ec.

ricardoV94 reviewed Jul 12, 2024

View reviewed changes

pytensor/link/jax/dispatch/tensor_basic.py Outdated Show resolved Hide resolved

pytensor/link/jax/dispatch/tensor_basic.py Outdated Show resolved Hide resolved

pytensor/link/jax/dispatch/tensor_basic.py Outdated Show resolved Hide resolved

Skip inline rewrite tests when mode == FAST_COMPILE

6487e61

ricardoV94 reviewed Jul 17, 2024

View reviewed changes

Incorporate review feedback

43474fd

ricardoV94 reviewed Jul 17, 2024

View reviewed changes

pytensor/tensor/rewriting/linalg.py Outdated Show resolved Hide resolved

pytensor/tensor/rewriting/ofg.py Outdated Show resolved Hide resolved

jessegrabowski added 2 commits July 17, 2024 15:17

Incorporate review feedback

6ef4084

Add is_zero_offset helper to Eye

04ddb46

ricardoV94 reviewed Jul 17, 2024

View reviewed changes

pytensor/tensor/elemwise.py Outdated Show resolved Hide resolved

pytensor/tensor/elemwise.py Show resolved Hide resolved

jessegrabowski added 3 commits July 18, 2024 09:19

Add is_left_expand_dims and is_right_expand_dims attributes to `D…

c038109

…imShuffle`

Seed test_local_lift_through_linalg test

19f2895

Fix failing diag_rewrite test

fcbccde

ricardoV94 approved these changes Jul 18, 2024

View reviewed changes

Revert symbolic offset

56a3ffe

jessegrabowski merged commit cac9feb into pymc-devs:main Jul 18, 2024
58 of 59 checks passed

jessegrabowski deleted the alloc-diag-ofg branch October 8, 2024 12:46

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add `OpFromGraph` wrapper around `alloc_diag` #915

Add `OpFromGraph` wrapper around `alloc_diag` #915

jessegrabowski commented Jul 10, 2024

jessegrabowski commented Jul 10, 2024

ricardoV94 commented Jul 10, 2024 •

edited

Loading

jessegrabowski commented Jul 11, 2024

ricardoV94 commented Jul 11, 2024

ricardoV94 commented Jul 11, 2024

ricardoV94 commented Jul 11, 2024

ricardoV94 left a comment

codecov bot commented Jul 12, 2024 •

edited

Loading

jessegrabowski commented Jul 12, 2024 •

edited

Loading

ricardoV94 commented Jul 12, 2024

jessegrabowski commented Jul 12, 2024

ricardoV94 commented Jul 12, 2024

ricardoV94 commented Jul 12, 2024

jessegrabowski commented Jul 12, 2024

jessegrabowski commented Jul 17, 2024

ricardoV94 commented Jul 17, 2024

jessegrabowski commented Jul 17, 2024

ricardoV94 commented Jul 17, 2024 •

edited

Loading

ricardoV94 left a comment

jessegrabowski commented Jul 18, 2024 •

edited

Loading

ricardoV94 commented Jul 18, 2024 •

edited

Loading

jessegrabowski commented Jul 18, 2024

ricardoV94 commented Jul 18, 2024 •

edited

Loading

ricardoV94 commented Jul 18, 2024

jessegrabowski commented Jul 18, 2024

ricardoV94 commented Jul 18, 2024 •

edited

Loading

ricardoV94 commented Jul 18, 2024

jessegrabowski commented Jul 18, 2024

jessegrabowski commented Jul 18, 2024 •

edited

Loading

ricardoV94 commented Jul 18, 2024 •

edited

Loading

Add OpFromGraph wrapper around alloc_diag #915

Add OpFromGraph wrapper around alloc_diag #915

Conversation

jessegrabowski commented Jul 10, 2024

Description

Related Issue

Checklist

Type of change

jessegrabowski commented Jul 10, 2024

ricardoV94 commented Jul 10, 2024 • edited Loading

jessegrabowski commented Jul 11, 2024

ricardoV94 commented Jul 11, 2024

ricardoV94 commented Jul 11, 2024

ricardoV94 commented Jul 11, 2024

ricardoV94 left a comment

Choose a reason for hiding this comment

codecov bot commented Jul 12, 2024 • edited Loading

Codecov Report

jessegrabowski commented Jul 12, 2024 • edited Loading

ricardoV94 commented Jul 12, 2024

jessegrabowski commented Jul 12, 2024

ricardoV94 commented Jul 12, 2024

ricardoV94 commented Jul 12, 2024

jessegrabowski commented Jul 12, 2024

jessegrabowski commented Jul 17, 2024

ricardoV94 commented Jul 17, 2024

jessegrabowski commented Jul 17, 2024

ricardoV94 commented Jul 17, 2024 • edited Loading

ricardoV94 left a comment

Choose a reason for hiding this comment

jessegrabowski commented Jul 18, 2024 • edited Loading

ricardoV94 commented Jul 18, 2024 • edited Loading

jessegrabowski commented Jul 18, 2024

ricardoV94 commented Jul 18, 2024 • edited Loading

ricardoV94 commented Jul 18, 2024

jessegrabowski commented Jul 18, 2024

ricardoV94 commented Jul 18, 2024 • edited Loading

ricardoV94 commented Jul 18, 2024

jessegrabowski commented Jul 18, 2024

jessegrabowski commented Jul 18, 2024 • edited Loading

ricardoV94 commented Jul 18, 2024 • edited Loading

Add `OpFromGraph` wrapper around `alloc_diag` #915

Add `OpFromGraph` wrapper around `alloc_diag` #915

ricardoV94 commented Jul 10, 2024 •

edited

Loading

codecov bot commented Jul 12, 2024 •

edited

Loading

jessegrabowski commented Jul 12, 2024 •

edited

Loading

ricardoV94 commented Jul 17, 2024 •

edited

Loading

jessegrabowski commented Jul 18, 2024 •

edited

Loading

ricardoV94 commented Jul 18, 2024 •

edited

Loading

ricardoV94 commented Jul 18, 2024 •

edited

Loading

ricardoV94 commented Jul 18, 2024 •

edited

Loading

jessegrabowski commented Jul 18, 2024 •

edited

Loading

ricardoV94 commented Jul 18, 2024 •

edited

Loading