Add default compiler options for v6e

axlearn/common/compiler_options.py

@@ -4,12 +4,12 @@
 # This module must not depend on any jax/axlearn modules so that
 # importing this module does not result in initializing jax.
 import re
-from typing import Any, Union
+from typing import Any, Dict, Union
 
 
 def default_xla_options(
     *, instance_type: str, num_slices: int, backend: str
-) -> dict[str, Union[str, bool]]:
+) -> dict[str, Union[str, bool, int]]:
     """Return the default flags for the given instance type and backend.
 
     These options can be passed to `jitted_fn.lower(...).compile(compiler_options=...)`
@@ -31,7 +31,7 @@ def default_xla_options(
     if backend != "tpu":
         raise NotImplementedError(backend)
     version = infer_tpu_version(infer_tpu_type(instance_type))
-    options = dict(
+    options: Dict[str, int | str | bool] = dict(
         xla_tpu_spmd_rng_bit_generator_unsafe=True,  # SPMD partition-aware RngBitGenerator.
         xla_tpu_enable_latency_hiding_scheduler="true",  # Try to schedule ops efficiently.
         xla_tpu_perform_spmd_cse_prevention="false",
@@ -44,6 +44,60 @@ def default_xla_options(
             xla_enable_async_all_gather="true",  # Allow async all-gather.
             xla_enable_async_collective_permute="true",  # Allow async collective permute.
         )
+    if version == "v6e":
+        options.update(
+            # Improved performance for v6e.
+            xla_tpu_scoped_vmem_limit_kib=98304,
+            xla_tpu_enable_async_collective_fusion="true",
+            xla_tpu_enable_async_collective_fusion_fuse_all_gather="true",
+            xla_tpu_enable_async_collective_fusion_multiple_steps="true",
+            xla_tpu_overlap_compute_collective_tc="true",
+            xla_enable_async_all_gather="true",
+            # Host offloading flags
+            xla_tpu_enable_all_experimental_scheduler_features="true",
+            # Flag to enable memory tracking scheduling. The default AUTO only enables
+            # it in some situations. Not needed if
+            # xla_tpu_enable_all_experimental_scheduler_features is set to true already.
+            xla_tpu_enable_scheduler_memory_pressure_tracking="true",
+            # Flag controlling the maximum number of overlapping host offloadings.
+            xla_tpu_host_transfer_overlap_limit=24,
+            # Flag to enable the aggressive removal of opt-barriers.
+            xla_tpu_aggressive_opt_barrier_removal="true",
+            # Flag to enable more aggressive scheduling for async ops, such as pushing
+            # the async start to the beginning of the loop body.
+            xla_lhs_prioritize_async_depth_over_stall="true",
+            # Flag to enable pipelining of cross-DCN all-gathers.
+            xla_tpu_enable_ag_backward_pipelining="true",
+            xla_should_allow_loop_variant_parameter_in_chain="true",
+            xla_should_add_loop_invariant_op_in_chain="true",
+            # Flag controlling the maximum number of overlapping cross-DCN send/recv.
+            xla_max_concurrent_host_send_recv=100,
+            # Flag controlling the HBM memory limit as a percentage of the total HBM size.
+            # Default value is 95. Can tune up or down to give more or less memory for the
+            # scheduler. The scheduler favors more on less memory usage when it's under
+            # memory pressure, instead of hiding latency by overlapping more computations
+            # and communications.
+            xla_tpu_scheduler_percent_shared_memory_limit=90,
+            # Flag controlling the number of times the scheduler is run if the scheduled
+            # peak memory usage exceeds the initial memory limit, by setting memory limit
+            # to 90% of the previous memory limit each time. Default value is 1. Sometimes
+            # when the scheduler thinks it goes out memory, it may not actually happen due
+            # to other factors controlled by other compiler passes, or the initial memory
+            # limit is already set too low. Cutting the memory limit to 90% of previous one
+            # though, may make the scheduler weighting too much on the memory usage instead
+            # of latency side.
+            xla_latency_hiding_scheduler_rerun=2,
+            xla_tpu_use_enhanced_launch_barrier="true",
+            # Sparsecore offloading for all reduce.
+            xla_sc_disable_megacore_partitioning="true",
+            xla_tpu_use_tc_device_shape_on_sc="true",
+            tpu_use_continuations="true",
+            xla_jf_crs_combiner_threshold_count=10,
+            xla_sc_enable_instruction_fusion="false",
+            xla_sc_disjoint_spmem="false",
+            xla_tpu_enable_sparse_core_collective_offload_all_reduce="true",
+        )
+        options["2a886c8_chip_config_name"] = "megachip_tccontrol"
     if num_slices > 1:
         # Support multiple TPU slices connected over a data center network.
         options.update(
@@ -55,11 +109,20 @@ def default_xla_options(
             xla_tpu_data_parallel_opt_different_sized_ops="true",
             # Group non-blocking DCN collectives into as few stages as possible.
             xla_tpu_enable_sunk_dcn_allreduce_done_with_host_reduction="true",
+            # Change to 16GB. The default is 4GB which is too small for larger models. This
+            # cause the step time to be double. You should increase this
+            # further if you see "Allocator failed to allocate". A feature
+            # to dynamically allocate may come later: b/380514965
+            megascale_grpc_premap_memory_bytes=17179869184,
         )
 
     # Validate options. Will never fail if this function is implemented correctly.
     for k, v in options.items():
-        assert v in [True, False, "true", "false"], (k, v)
+        try:
+            int(v)
+            continue
+        except ValueError:
+            assert v in [True, False, "true", "false", "megachip_tccontrol"], (k, v)
 
     return options