Refactor Megablox Ops to use public Tokamax API

PiperOrigin-RevId: 900763560

Author: CaptainO5

src/maxtext/configs/base.yml

@@ -223,16 +223,6 @@ wo_tile_drhs_batch_seq: 512
 wo_tile_drhs_embed_dim: 1024
 wo_tile_drhs_mlp_dim: 1024
 
-wi_tile_fwd_buffer_count: 2
-wi_tile_dlhs_buffer_count: 2
-wi_tile_drhs_buffer_count: 2
-wo_tile_fwd_buffer_count: 2
-wo_tile_dlhs_buffer_count: 2
-wo_tile_drhs_buffer_count: 2
-
-wi_combine_scopes: False
-wo_combine_scopes: False
-
 merge_gating_gmm: False
 
 norm_topk_prob: false # boolean to enable the top-k probability normalization. qwen3-specific normalization of router weights.

src/maxtext/configs/types.py

@@ -734,16 +734,6 @@ class MoEKernels(BaseModel):
   wo_tile_drhs_embed_dim: int = Field(1024, description="bwd pass drhs tiling dimension for embedding in GMM for wo.")
   wo_tile_drhs_mlp_dim: int = Field(1024, description="bwd pass drhs tiling dimension for MLP in GMM for wo.")
 
-  wi_tile_fwd_buffer_count: int = Field(2, description="forward pass tiling buffer count in GMM for wi.")
-  wi_tile_dlhs_buffer_count: int = Field(2, description="bwd pass dlhs tiling buffer count in GMM for wi.")
-  wi_tile_drhs_buffer_count: int = Field(2, description="bwd pass drhs tiling buffer count in GMM for wi.")
-  wo_tile_fwd_buffer_count: int = Field(2, description="forward pass tiling buffer count in GMM for wo.")
-  wo_tile_dlhs_buffer_count: int = Field(2, description="bwd pass dlhs tiling buffer count in GMM for wo.")
-  wo_tile_drhs_buffer_count: int = Field(2, description="bwd pass drhs tiling buffer count in GMM for wo.")
-
-  wi_combine_scopes: bool = Field(False, description="whether to use combine_scopes features for tgmm for wi.")
-  wo_combine_scopes: bool = Field(False, description="whether to use combine_scopes features for tgmm for wo.")
-
   merge_gating_gmm: bool = Field(False, description="whether to merge the two gating gmm kernels into one.")
 
 

src/maxtext/kernels/megablox/ops.py

@@ -16,23 +16,40 @@
 
 # pylint: disable=too-many-positional-arguments
 
-import functools
 import dataclasses
-from typing import Literal, List, Tuple
+import functools
+from typing import List, Literal, Tuple
 import jax
 import jax.numpy as jnp
 from maxtext.kernels.megablox import backend
-from tokamax._src.ops.ragged_dot import pallas_mosaic_tpu_kernel as tokamax_backend
 import qwix
 import qwix.pallas as qpl
+import tokamax
+
+
+DRHS_RAGGED_DOT_DIM_NUMS = jax.lax.RaggedDotDimensionNumbers(
+    dot_dimension_numbers=(([0], [0]), ([], [])),
+    lhs_ragged_dimensions=[0],
+    rhs_group_dimensions=[],
+)
 
 
 def gmm(
     lhs: jnp.ndarray,
     rhs: jnp.ndarray,
     group_sizes: jnp.ndarray,
     preferred_element_type: jnp.dtype = jnp.float32,
-    tiling: tuple[int, int, int, int, int, int, int, int, int] = (128, 128, 128, 128, 128, 128, 128, 128, 128),
+    tiling: tuple[int, int, int, int, int, int, int, int, int] = (
+        128,
+        128,
+        128,
+        128,
+        128,
+        128,
+        128,
+        128,
+        128,
+    ),
     group_offset: jnp.ndarray | None = None,
     existing_out: jnp.ndarray | None = None,
     transpose_rhs: bool = False,
@@ -42,8 +59,6 @@ def gmm(
     use_qwix_quantization: bool = False,
     use_tokamax_backend: bool = False,
     weight_gather_axes: List[Tuple[str, int]] | None = None,
-    input_buffer_count: tuple[int, int, int] = (2, 2, 2),
-    combine_scopes: bool = False,
     # TODO(amandaliang): get rid of the qwix_rule in favor of Qwix's interception feature
     qwix_rule: qwix.QtRule | None = None,
 ):
@@ -65,16 +80,14 @@ def gmm(
       )
 
   gmm_fwd_bwd = lambda *args: _gmm_fwd(*args)[0]  # pylint: disable=C3001
-  gmm_fwd_bwd = jax.custom_vjp(gmm_fwd_bwd, nondiff_argnums=(3, 4, 5, 6, 9, 10, 11, 12, 13))
+  gmm_fwd_bwd = jax.custom_vjp(gmm_fwd_bwd, nondiff_argnums=(3, 4, 7, 8, 9, 10, 11))
   gmm_fwd_bwd.defvjp(_gmm_fwd, functools.partial(_gmm_bwd, lhs.dtype, rhs.dtype))
   return gmm_fwd_bwd(
       lhs,
       rhs,
       group_sizes,
       preferred_element_type,
       tiling,
-      input_buffer_count,
-      combine_scopes,
       group_offset,
       existing_out,
       transpose_rhs,
@@ -90,9 +103,17 @@ def _gmm_fwd(
     rhs: jnp.ndarray,
     group_sizes: jnp.ndarray,
     preferred_element_type: jnp.dtype = jnp.float32,
-    tiling: tuple[int, int, int, int, int, int, int, int, int] = (128, 128, 128, 128, 128, 128, 128, 128, 128),
-    input_buffer_count: tuple[int, int, int] = (2, 2, 2),
-    combine_scopes: bool = False,
+    tiling: tuple[int, int, int, int, int, int, int, int, int] = (
+        128,
+        128,
+        128,
+        128,
+        128,
+        128,
+        128,
+        128,
+        128,
+    ),
     group_offset: jnp.ndarray | None = None,
     existing_out: jnp.ndarray | None = None,
     transpose_rhs: bool = False,
@@ -136,17 +157,18 @@ def _gmm_fwd(
           for axis_name, axis_idx in weight_gather_axes:
             rhs_qvalue = jax.lax.all_gather(rhs.qvalue, axis_name, axis=axis_idx, tiled=True)
             rhs = dataclasses.replace(rhs, qvalue=rhs_qvalue)
-    out = tokamax_backend.gmm(
+    # Handle transpose_rhs manually as ragged_dot assumes (G, K, N)
+    if transpose_rhs:
+      rhs = rhs.swapaxes(1, 2)
+
+    out = tokamax.ragged_dot(
         lhs=lhs,
         rhs=rhs,
         group_sizes=group_sizes,
         precision=jax.lax.Precision.DEFAULT,
-        out_dtype=preferred_element_type,
-        tiling=tiling[:3],
+        preferred_element_type=preferred_element_type,
         group_offset=group_offset,
-        transpose_rhs=transpose_rhs,
-        interpret=interpret,
-        input_buffer_count=input_buffer_count[0],
+        implementation="mosaic",
     )
   else:
     out = backend.gmm(
@@ -168,8 +190,6 @@ def _gmm_bwd(
     rhs_dtype: jax.typing.DTypeLike,
     preferred_element_type: jnp.dtype,
     tiling: tuple[int, int, int, int, int, int, int, int, int],
-    input_buffer_count: tuple[int, int, int],
-    combine_scopes: bool,
     transpose_rhs: bool,
     interpret: bool,
     quantization_rule: qwix.QtRule | None,
@@ -224,30 +244,29 @@ def _gmm_bwd(
         calibration_method=quantization_rule.bwd_calibration_method,
     )
   if use_tokamax_backend:
-    dlhs = tokamax_backend.gmm(
+    # Handle transpose_rhs manually
+    dlhs_rhs = rhs
+    if not transpose_rhs:
+      dlhs_rhs = dlhs_rhs.swapaxes(1, 2)
+
+    dlhs = tokamax.ragged_dot(
         lhs=dlhs_dout,
-        rhs=rhs,
+        rhs=dlhs_rhs,
         group_sizes=group_sizes,
         precision=jax.lax.Precision.DEFAULT,
-        out_dtype=lhs_dtype,
-        tiling=tiling[3:6],
+        preferred_element_type=lhs_dtype,
         group_offset=group_offset,
-        transpose_rhs=not transpose_rhs,
-        interpret=interpret,
-        input_buffer_count=input_buffer_count[1],
+        implementation="mosaic",
     )
-    drhs = tokamax_backend.tgmm(
-        lhs=lhs.swapaxes(0, 1),
+    drhs = tokamax.ragged_dot_general(
+        lhs=lhs,
         rhs=drhs_dout,
         group_sizes=group_sizes,
+        ragged_dot_dimension_numbers=DRHS_RAGGED_DOT_DIM_NUMS,
         precision=jax.lax.Precision.DEFAULT,
-        out_dtype=rhs_dtype,
-        tiling=tiling[-3:],
+        preferred_element_type=rhs_dtype,
         group_offset=group_offset,
-        num_actual_groups=num_actual_groups,
-        interpret=interpret,
-        input_buffer_count=input_buffer_count[2],
-        combine_scopes=combine_scopes,
+        implementation="mosaic",
     )
     if quantization_rule and quantization_rule.bwd_qtype and weight_gather_axes:
       # Scatter back in reverse order of gather

src/maxtext/layers/moe.py

@@ -965,9 +965,7 @@ def get_quantization_dtypes():
         rhs_quantize_dtype = quant_dg.fwd.dg_quantizer.rhs.numerics.get_dtype()
       return lhs_quantize_dtype, rhs_quantize_dtype
 
-    def gmm(
-        inputs, kernel, tiling, group_sizes, expert_assignments, weight_gather_axes, input_buffer_count, combine_scopes
-    ):
+    def gmm(inputs, kernel, tiling, group_sizes, expert_assignments, weight_gather_axes):
       if inputs.shape[0] != expert_assignments.shape[0]:
         raise ValueError("The number of input tokens must match the number of expert assignments!")
 
@@ -993,8 +991,6 @@ def gmm(
               use_qwix_quantization=self.config.use_qwix_quantization,
               use_tokamax_backend=self.config.use_tokamax_gmm,
               weight_gather_axes=weight_gather_axes,
-              input_buffer_count=input_buffer_count,
-              combine_scopes=combine_scopes,
           )
         else:  # tokamax (unquantized)
           output = tokamax.ragged_dot(
@@ -1250,26 +1246,12 @@ def get_active_sharding_axes(pspec_dim_axes, tensor_dim_index):
           self.config.wo_tile_drhs_embed_dim,
           self.config.wo_tile_drhs_mlp_dim,
       )
-      wi_input_buffer_count = (
-          self.config.wi_tile_fwd_buffer_count,
-          self.config.wi_tile_dlhs_buffer_count,
-          self.config.wi_tile_drhs_buffer_count,
-      )
-      wo_input_buffer_count = (
-          self.config.wo_tile_fwd_buffer_count,
-          self.config.wo_tile_dlhs_buffer_count,
-          self.config.wo_tile_drhs_buffer_count,
-      )
 
-      wi_combine_scopes = self.config.wi_combine_scopes
-      wo_combine_scopes = self.config.wo_combine_scopes
       layer_w0 = gmm_fn(
           x,
           w0,
           tiling=wi_tile_size,
           weight_gather_axes=wi_gather_axes,
-          input_buffer_count=wi_input_buffer_count,
-          combine_scopes=wi_combine_scopes,
       )
       if self.get_tensor_transpose_parallelism_size() > 1:
         layer_w0 = jax.lax.psum(layer_w0, "tensor_transpose")
@@ -1282,8 +1264,6 @@ def get_active_sharding_axes(pspec_dim_axes, tensor_dim_index):
           w1,
           tiling=wi_tile_size,
           weight_gather_axes=wi_gather_axes,
-          input_buffer_count=wi_input_buffer_count,
-          combine_scopes=wi_combine_scopes,
       )
       if self.get_tensor_transpose_parallelism_size() > 1:
         layer_w1 = jax.lax.psum(layer_w1, "tensor_transpose")
@@ -1297,8 +1277,6 @@ def get_active_sharding_axes(pspec_dim_axes, tensor_dim_index):
           wo,
           tiling=wo_tile_size,
           weight_gather_axes=wo_gather_axes,
-          input_buffer_count=wo_input_buffer_count,
-          combine_scopes=wo_combine_scopes,
       )
       if self.get_tensor_parallelism_size() > 1:
         intermediate_output = jax.lax.psum_scatter(

src/maxtext/models/deepseek_batchsplit_fp8.py

@@ -948,8 +948,6 @@ def gmm(
       group_sizes,
       preferred_element_type,
       weight_gather_axes,
-      input_buffer_count,
-      combine_scopes,
   ):
     if config.use_qwix_quantization:
       output = megablox.gmm(
@@ -961,8 +959,6 @@ def gmm(
           use_qwix_quantization=config.use_qwix_quantization,
           use_tokamax_backend=config.use_tokamax_gmm,
           weight_gather_axes=weight_gather_axes,
-          input_buffer_count=input_buffer_count,
-          combine_scopes=combine_scopes,
           qwix_rule=quantizations.get_fp8_full_qwix_rule(config),
       )
     else:
@@ -1002,19 +998,7 @@ def gmm(
       config.wo_tile_drhs_embed_dim,
       config.wo_tile_drhs_mlp_dim,
   )
-  wi_input_buffer_count = (
-      config.wi_tile_fwd_buffer_count,
-      config.wi_tile_dlhs_buffer_count,
-      config.wi_tile_drhs_buffer_count,
-  )
-  wo_input_buffer_count = (
-      config.wo_tile_fwd_buffer_count,
-      config.wo_tile_dlhs_buffer_count,
-      config.wo_tile_drhs_buffer_count,
-  )
 
-  wi_combine_scopes = config.wi_combine_scopes
-  wo_combine_scopes = config.wo_combine_scopes
   if config.use_qwix_quantization:
     gating_pspec, linear_pspec = moe_lib.get_batchsplit_init_kernel_axes()
     w0_pspec = nn.logical_to_mesh_axes(gating_pspec)
@@ -1043,8 +1027,6 @@ def get_active_sharding_axes(pspec_dim_axes, tensor_dim_index):
         w01,
         tiling=wi_tile_size,
         weight_gather_axes=wi_gather_axes,
-        input_buffer_count=wi_input_buffer_count,
-        combine_scopes=wi_combine_scopes,
     )
     layer_w0, layer_w1 = jnp.split(layer_w01, 2, axis=-1)
   else:
@@ -1053,16 +1035,12 @@ def get_active_sharding_axes(pspec_dim_axes, tensor_dim_index):
         w0,
         tiling=wi_tile_size,
         weight_gather_axes=wi_gather_axes,
-        input_buffer_count=wi_input_buffer_count,
-        combine_scopes=wi_combine_scopes,
     )
     layer_w1 = gmm_fn(
         x,
         w1,
         tiling=wi_tile_size,
         weight_gather_axes=wi_gather_axes,
-        input_buffer_count=wi_input_buffer_count,
-        combine_scopes=wi_combine_scopes,
     )
   layer_w0 = jax.ad_checkpoint.checkpoint_name(layer_w0, "mlpwi_0")
   layer_w1 = jax.ad_checkpoint.checkpoint_name(layer_w1, "mlpwi_1")
@@ -1073,8 +1051,6 @@ def get_active_sharding_axes(pspec_dim_axes, tensor_dim_index):
       wo,
       tiling=wo_tile_size,
       weight_gather_axes=wo_gather_axes,
-      input_buffer_count=wo_input_buffer_count,
-      combine_scopes=wo_combine_scopes,
   )
   return layer_wo