Return DShapeArray objects in measure primitives when possible #1170
Conversation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #1170 +/- ##
==========================================
+ Coverage 97.89% 97.96% +0.07%
==========================================
Files 76 76
Lines 10879 10885 +6
Branches 1289 1292 +3
==========================================
+ Hits 10650 10664 +14
+ Misses 178 174 -4
+ Partials 51 47 -4 ☔ View full report in Codecov by Sentry. |
rauletorresc
changed the title
| shots, shape_value = jaxpr.eqns[1].params.values() | ||
|
|
||
| assert jaxpr.eqns[1].primitive == probs_p | ||
| assert isinstance(shape_value[0], DynamicJaxprTracer) | ||
| assert shots == 5 |
There was a problem hiding this comment.
Do you know what changed that means there are no escaped tracer issues compared to your previous test?
There was a problem hiding this comment.
Note the jaxpr for the example here is different than what JAX would generate, and I'm wondering if that isn't going to lead to issues (like escaped tracers).
Our jaxpr, contains tracer in the jaxpr equation:
{ lambda ; a:i64[]. let
b:AbstractObs(num_qubits=0,primitive=compbasis) = compbasis
c:f64[a] d:i64[a] = counts[
shape=(Traced<ShapedArray(int64[], weak_type=True)>with<DynamicJaxprTrace(level=1/0)>,)
shots=5
] b
in (c, d) }JAX's jaxpr for similar example (jnp.ones((dim_0, 5))), does not insert the tracer as constant data into the jaxpr equation:
{ lambda ; a:i64[]. let
b:f64[a,5] = broadcast_in_dim[broadcast_dimensions=() shape=(None, 5)] 1.0 a
in (b,) }Instead we find None in the place where the tracer would be.
There was a problem hiding this comment.
I managed to remove shots from the jaxpr by making it a traceable argument. However, shape is still a keyword argument, that's why we still see the shots tracer as constant data inside shape. Once we support shape as traceable, I guess we will see the None keyword instead.
| def f(dim_0): | ||
| obs = compbasis_p.bind() | ||
| return state_p.bind(obs, shots=5, shape=(dim_0,)) | ||
|
|
||
| jaxpr = jax.make_jaxpr(f)(1) | ||
| shots, shape_value = jaxpr.eqns[1].params.values() |
There was a problem hiding this comment.
I think we still need to cover more of the pipeline in our tests (ignoring the MLIR core for now, which can be addressed in a different PR).
From what I can see we test the bind function and abstract evaluation of jaxpr primitives, but that leaves out most of the code that constitutes the frontend. The first step might be to identify which functions could be affected by this change, from the start of tracing until jax spits out the mlir, and then ensure those functions are tested for the dynamic case. Or alternatively develop tests that cover roughly that entire span.
|
I decided to tackle the MLIR changes in the same PR |
Context: Measure primitives only return
ShapedArrayobjects, which forces variables like the number of shots or number of wires to be known at compilation time. With the support forDShapedArrayobjects, this constraint can be removed.Description of the Change: Return
DShapedArrayobjects when possible.Benefits: Eliminating the need to know the number of shots/wires statically is a step towards allowing us to compile programs that don't have a fixed number of qubits.
[sc-74736]