[jaxprs] The next step in ClosedJaxpr simplification

Now after #29882 non-scalar constants appear as `core.Literal`
in `Jaxpr.eqns.invars`, and `ClosedJaxpr` are mostly empty of `consts`.
Before this change the lowering would serialize the constant as
a HLO constant. There are several problems with this:

  * if the constant is on a device, it gets copied to the host and
  is embedded in the lowered HLO. This increases memory usage,
  and if the constant was sharded on multiple devices it will
  now become replicated.
  * if a constant appears in multiple places in the Jaxprs, today
  we will make copies.
  * if the constants are inlined in the HLO the XLA constant folding
  starts to constant-fold them, and we see warnings of very slow
  constant folding.
  * the constant folding changes the numerical behavior compared
  to current lowering strategy.

With this change we hoist the constants out of the Jaxpr (the
new function `jaxpr_hoisted_consts`, we pass them as argument to
the top-level function and we thread them to other functions
that need them. We add new arguments to functions to represent
the hoisted constants.

This change has no effect, unless JAX_USE_SIMPLIFIED_JAXPR_CONSTANTS=True.

This is still an experiment.

Author: gnecula

jax/BUILD

@@ -1235,6 +1235,7 @@ pytype_strict_library(
         ":sharding_impls",
         ":source_info_util",
         ":state_types",
+        ":typing",
         ":util",
         ":xla",
         ":xla_bridge",

jax/_src/ad_checkpoint.py

@@ -783,7 +783,7 @@ def _remat_translation_using_opt_barrier(*args, jaxpr: core.Jaxpr):
 
 
 def _remat_lowering(
-    ctx,
+    ctx: mlir.LoweringRuleContext,
     *args,
     jaxpr: core.Jaxpr,
     prevent_cse: bool,
@@ -801,7 +801,8 @@ def _remat_lowering(
     jaxpr_args = args
   outs, tokens_out = mlir.jaxpr_subcomp(
       ctx.module_context, jaxpr, ctx.name_stack.extend('checkpoint'),
-      ctx.tokens_in, (), *jaxpr_args, dim_var_values=ctx.dim_var_values)
+      ctx.tokens_in, (), *jaxpr_args, dim_var_values=ctx.dim_var_values,
+      const_lowering=ctx.const_lowering)
   ctx.set_tokens_out(tokens_out)
   return outs
 

jax/_src/api.py

@@ -1690,8 +1690,8 @@ def cache_miss(*args, **kwargs):
     with core.take_current_trace() as trace:
       try:
         if isinstance(trace, core.EvalTrace):
-          execute = pxla.xla_pmap_impl_lazy(p.flat_fun, *p.flat_args, **params)
-          out = execute(*p.flat_args)
+          execute, hoisted_consts = pxla.xla_pmap_impl_lazy(p.flat_fun, *p.flat_args, **params)
+          out = execute(*hoisted_consts, *p.flat_args)
         else:
           out = pxla.xla_pmap_p.bind_with_trace(trace, (p.flat_fun, *p.flat_args), params)
       except api_util.InternalFloatingPointError as e:

jax/_src/core.py

@@ -59,7 +59,7 @@
 from jax._src.lib import jax_jit
 from jax._src.lib import xla_client
 from jax._src import traceback_util
-from jax._src.typing import Array, DimSize, Shape
+from jax._src.typing import Array, ArrayLike, DimSize, Shape
 from jax._src import typing
 from jax._src import xla_metadata_lib
 
@@ -548,6 +548,33 @@ def is_literalable(x: Any) -> bool:
       return (not np.shape(x) or config.use_simplified_jaxpr_constants.value)
   return False
 
+@partial(weakref_lru_cache, trace_context_in_key=False)
+def jaxpr_hoisted_consts(jaxpr: Jaxpr) -> list[ArrayLike]:
+  # The non-scalar constants in core.Literal, in the entire Jaxpr,
+  # uniquified by id. These will be hoisted as arguments to the functions
+  # in which they appear.
+  if not config.use_simplified_jaxpr_constants.value:
+    return []
+  consts_by_id: dict[int, Any] = {}
+  for v in jaxpr.outvars:
+    if type(v) is Literal and np.shape(v.val):  # type: ignore
+      consts_by_id[id(v)] = v.val  # type: ignore
+
+  for eqn in jaxpr.eqns:
+    for v in eqn.invars:
+      if type(v) is Literal and np.shape(v.val):  # type: ignore
+        consts_by_id[id(v)] = v.val  # type: ignore
+    consts_by_id.update({id(v): v
+                         for v in eqn_params_hoisted_consts(eqn.params)})
+  return list(consts_by_id.values())
+
+def eqn_params_hoisted_consts(params) -> list[ArrayLike]:
+  consts_by_id: dict[int, Any] = {}
+  for j in jaxprs_in_params(params):
+      consts_by_id.update({id(v): v for v in jaxpr_hoisted_consts(j)})
+
+  return list(consts_by_id.values())
+
 Atom = Union[Var, Literal]
 
 class Primitive:

jax/_src/custom_derivatives.py

@@ -437,11 +437,13 @@ def _custom_jvp_call_typecheck(_, *in_avals, call_jaxpr, jvp_jaxpr_fun,
   return call_jaxpr.out_avals, call_jaxpr.effects
 core.custom_typechecks[custom_jvp_call_p] = _custom_jvp_call_typecheck
 
-def _custom_jvp_vjp_call_lowering(ctx, *args, call_jaxpr, **_):
+def _custom_jvp_vjp_call_lowering(ctx: mlir.LoweringRuleContext, *args,
+                                  call_jaxpr: core.ClosedJaxpr, **_):
   consts = mlir._ir_consts(call_jaxpr.consts)
   out, tokens = mlir.jaxpr_subcomp(ctx.module_context, call_jaxpr.jaxpr,
                                    ctx.name_stack, ctx.tokens_in, consts,
-                                   *args, dim_var_values=ctx.dim_var_values)
+                                   *args, dim_var_values=ctx.dim_var_values,
+                                   const_lowering=ctx.const_lowering)
   ctx.set_tokens_out(tokens)
   return out
 mlir.register_lowering(custom_jvp_call_p, _custom_jvp_vjp_call_lowering)

jax/_src/export/_export.py

@@ -617,6 +617,7 @@ def do_export(*args_specs, **kwargs_specs) -> Exported:
         _private_parameters=mlir.LoweringParameters(
             override_lowering_rules=override_lowering_rules,
             for_export=True,
+            hoist_constants_as_args=False,
             export_ignore_forward_compatibility=config.export_ignore_forward_compatibility.value))
     return _export_lowered(
         lowered, traced.jaxpr, traced.fun_name,
@@ -954,15 +955,16 @@ def is_token(typ, attrs):
           host_callbacks=[], module=wrapped_module, context=context,
           lowering_parameters=mlir.LoweringParameters(
               global_constant_computation=True,
-              for_export=True,
+              for_export=True, hoist_constants_as_args=False,
               export_ignore_forward_compatibility=config.export_ignore_forward_compatibility.value,
           ))
       ctx = mlir.LoweringRuleContext(
         module_context=module_context,
         name_stack=source_info_util.new_name_stack(), traceback=None,
         primitive=None,
         avals_in=args_avals_flat, avals_out=None,
-        tokens_in=mlir.TokenSet(), tokens_out=None)
+        tokens_in=mlir.TokenSet(), tokens_out=None,
+        const_lowering={})
       # We compute dim_values from the array arguments.
       new_main_op_array_args = new_main_op.arguments[-nr_array_args:]
       if shape_poly.all_dim_vars(args_avals_flat):

jax/_src/frozen_dict.py

@@ -47,3 +47,6 @@ def __iter__(self) -> Iterator[K]:
 
   def __len__(self) -> int:
     return len(self._d)
+
+  def get(self, key: K) -> V | None:  # type: ignore
+    return self._d.get(key, None)