Refactor llama / mixtral / grok for shared features (#267)

Many of these features can toggle between depending on architecture.
Replumbing the configurations separately allows better reuse and
understanding of how models vary between eachother.

grok uses a softcap, plumbing a value enables `sc * tanh( v / sc)` grok
has some hardcoded values that have better representations, e.g.
`sqrt(6144)` and `sqrt(3)`.

output normalization is optional but used by mixtral. Presence of the
tensor is sufficient for performing the normalization.

sharktank/sharktank/export_layer/export_moe.py

@@ -5,9 +5,12 @@
 # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
 import torch
+import torch.nn.functional as F
+
 from iree.turbine.aot import *
+
 from sharktank.models.llama.testing import make_moe_block_theta, make_rand_torch
-from sharktank.layers.mixture_of_experts_block import PreGatherMoeBlock
+from sharktank.layers.mixture_of_experts_block import MoeBlock
 from ..utils import cli
 
 
@@ -37,21 +40,21 @@ def main():
         action="store_true",
     )
     parser.add_argument(
-        "--use-grok",
-        help="Enable to export Grok model's version of MOE block",
+        "--use-gelu",
+        help="Enable to use gelu for moe activation",
         action="store_true",
     )
 
     args = cli.parse(parser)
 
     bs = args.batch_size
 
-    model = PreGatherMoeBlock(
+    model = MoeBlock(
         theta=make_moe_block_theta()("blk.0"),
         expert_count=8,
         expert_used_count=2,
         rms_epsilon=1e-5,
-        use_grok=args.use_grok,
+        moe_activation=F.gelu if args.use_gelu else F.silu,
     )
     fxb = FxProgramsBuilder(model)
     input = make_rand_torch((bs, 32, 6144))

sharktank/sharktank/layers/__init__.py

@@ -16,6 +16,6 @@
 from .paged_llama_attention_block import PagedLlamaAttentionBlock
 from .ffn_block import FFN
 from .ffn_moe_block import FFNMOE
-from .mixture_of_experts_block import SparseMoeBlock, PreGatherMoeBlock
+from .mixture_of_experts_block import MoeBlock
 
 from .configs import *

sharktank/sharktank/layers/ffn_moe_block.py

@@ -12,7 +12,7 @@
 from .base import ThetaLayer
 from .linear import LinearLayer
 from ..types import Theta, DefaultPrimitiveTensor
-from ..ops import einsum_2args
+from ..ops import einsum_2args, elementwise
 
 __all__ = [
     "FFNMOE",
@@ -24,15 +24,15 @@ class PreGatherFFNMOE(ThetaLayer):
     def __init__(
         self,
         theta: Theta,
-        use_grok: bool = False,
+        activation=F.silu,
     ):
 
         super().__init__(theta)
-        self.use_grok = use_grok
 
         self.ffn_gate = theta.tensor("ffn_gate_exps", "weight")
         self.ffn_up = theta.tensor("ffn_up_exps", "weight")
         self.ffn_down = theta.tensor("ffn_down_exps", "weight")
+        self.activation = activation
 
     def pre_matmul_gather(self, inputs, weights, experts, einstring="mk,menk->men"):
         inputs = inputs[:, :]
@@ -63,10 +63,8 @@ def forward(
         experts: torch.Tensor,
         expert_gate: torch.Tensor,
     ):
-        if self.use_grok:
-            ffn_gate = F.gelu(self.pre_matmul_gather(h, self.ffn_gate, experts))
-        else:
-            ffn_gate = F.silu(self.pre_matmul_gather(h, self.ffn_gate, experts))
+        ffn_gate = self.pre_matmul_gather(h, self.ffn_gate, experts)
+        ffn_gate = elementwise(self.activation, ffn_gate)
 
         ffn_up = self.pre_matmul_gather(h, self.ffn_up, experts)
         ffn_down = self.pre_matmul_gather(

sharktank/sharktank/layers/llama_attention_block.py

@@ -6,8 +6,6 @@
 
 from typing import Optional
 
-import math
-
 import torch
 import torch.nn.functional as F
 
@@ -110,7 +108,9 @@ def repeat_kv(x: torch.Tensor) -> torch.Tensor:
         values = values.transpose(1, 2)
 
         # Flash attention.
-        attn_weights = torch.matmul(xq, keys.transpose(2, 3)) / math.sqrt(self.head_dim)
+        attn_weights = torch.matmul(xq, keys.transpose(2, 3)) / torch.sqrt(
+            self.head_dim
+        )
 
         # Apply attention mask.
         if attention_mask is not None:

sharktank/sharktank/layers/mixture_of_experts_block.py

@@ -16,12 +16,11 @@
 from .ffn_moe_block import FFNMOE, PreGatherFFNMOE
 
 __all__ = [
-    "SparseMoeBlock",
-    "PreGatherMoeBlock",
+    "MoeBlock",
 ]
 
 
-class SparseMoeBlock(ThetaLayer):
+class MoeBlock(ThetaLayer):
     """
     This implementation considers MoE operations as block-sparse
     operations to support imbalanced token assignments to experts.
@@ -35,108 +34,12 @@ def __init__(
         expert_count: int,
         expert_used_count: int,
         rms_epsilon: float,
-    ):
-        super().__init__(theta)
-
-        # Add router gate
-        self.add_module("ffn_gate_inp", LinearLayer(theta("ffn_gate_inp")))
-
-        # Add FFN norm
-        self.add_module(
-            "ffn_norm", RMSNormLayer(theta("ffn_norm"), epsilon=rms_epsilon)
-        )
-
-        # Add FFN output norm
-        self.add_module(
-            "layer_output_norm",
-            RMSNormLayer(theta("layer_output_norm"), epsilon=rms_epsilon),
-        )
-
-        # Add expert_count x FFN
-        self.experts = nn.ModuleList(
-            [FFNMOE(theta, expert_idx=i) for i in range(expert_count)]
-        )
-
-        self.expert_count = expert_count
-        self.expert_used_count = expert_used_count
-
-    def forward(
-        self,
-        h: torch.Tensor,
-    ):
-        ffn_input = self.ffn_norm(h)
-        batch_size, sequence_length, feature_dim = ffn_input.shape
-        ffn_input = ffn_input.view(-1, feature_dim)
-
-        # For each token, the router calculates the router weights for all experts
-        # router_logits: (batch_size * sequence_length, expert_count)
-        router_logits = self.ffn_gate_inp(ffn_input)
-        router_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
-
-        # Select top k experts from router weights
-        router_weights, top_k_experts = torch.topk(
-            router_weights, self.expert_used_count, dim=-1
-        )
-        router_weights /= router_weights.sum(dim=-1, keepdim=True)
-        router_weights = router_weights.to(ffn_input.dtype)
-
-        moe_output = torch.zeros(
-            (batch_size * sequence_length, feature_dim), dtype=ffn_input.dtype
-        )
-
-        # Create an expert mask by one hot encoding the selected top k experts
-        # used to index which expert is to be invoked for each token
-        # expert_mask: (expert_count, expert_used_count, sequence_length)
-        expert_mask = F.one_hot(top_k_experts, num_classes=self.expert_count).permute(
-            2, 1, 0
-        )
-
-        # Iterate over all experts in the model
-        for expert_idx in range(self.expert_count):
-            expert_layer = self.experts[expert_idx]
-            top_k_expert_idx, token_idx = torch.where(expert_mask[expert_idx])
-
-            # Given the hidden states, index the tokens assigned to this expert
-            # and calculate the current expert's hidden state and weigh the
-            # output expert hidden states by the router weights, based on the
-            # appropriate tokens
-            current_expert_tokens = ffn_input[None, token_idx]
-
-            current_expert = (
-                expert_layer(current_expert_tokens)
-                * router_weights[token_idx, top_k_expert_idx, None]
-            )
-
-            current_expert = current_expert.reshape(-1, feature_dim)
-
-            moe_output.index_add_(0, token_idx, current_expert.to(ffn_input.dtype))
-        moe_output = moe_output.reshape(batch_size, sequence_length, feature_dim)
-
-        moe_output = self.layer_output_norm(moe_output)
-        return h + moe_output
-
-
-class PreGatherMoeBlock(ThetaLayer):
-    """
-    This implementation considers MoE operations as block-sparse
-    operations to support imbalanced token assignments to experts.
-    This enables the MoE to operate at a faster rate and in full capacity without any dropped tokens
-    (or reduced performance).
-    """
-
-    def __init__(
-        self,
-        theta: Theta,
-        expert_count: int,
-        expert_used_count: int,
-        rms_epsilon: float,
-        use_grok: Optional[bool] = False,
+        moe_activation=F.silu,
     ):
         super().__init__(theta)
 
         self.expert_count = expert_count
         self.expert_used_count = expert_used_count
-        self.use_grok = use_grok
 
         # Add router gate
         self.add_module("ffn_gate_inp", LinearLayer(theta("ffn_gate_inp")))
@@ -146,15 +49,17 @@ def __init__(
             "ffn_norm", RMSNormLayer(theta("ffn_norm"), epsilon=rms_epsilon)
         )
 
-        # Add FFN output norm layer for Grok
-        if self.use_grok:
+        # Add optional FFN output norm layer
+        if theta.optional_tensor("layer_output_norm") is not None:
             self.add_module(
                 "layer_output_norm",
                 RMSNormLayer(theta("layer_output_norm"), epsilon=rms_epsilon),
             )
+        else:
+            self.add_module("layer_output_norm", torch.nn.Identity())
 
         # Add expert_count x FFN
-        self.experts = PreGatherFFNMOE(theta, use_grok=self.use_grok)
+        self.experts = PreGatherFFNMOE(theta, activation=moe_activation)
 
     def forward(
         self,
@@ -180,7 +85,6 @@ def forward(
         moe_output = self.experts(ffn_input, top_k_experts, expert_gate)
         moe_output = moe_output.reshape(batch_size, sequence_length, feature_dim)
 
-        if self.use_grok:
-            moe_output = self.layer_output_norm(moe_output)
+        moe_output = self.layer_output_norm(moe_output)
 
         return h + moe_output

sharktank/sharktank/layers/paged_llama_attention_block.py

@@ -37,7 +37,8 @@ def __init__(
         head_dim: int,
         head_count_kv: int,
         rms_epsilon: float,
-        use_grok: Optional[bool] = False,
+        attention_scale: Optional[float] = None,
+        softcap: Optional[float] = None,
     ):
         super().__init__(theta)
 
@@ -46,7 +47,8 @@ def __init__(
         self.head_count = head_count
         self.head_dim = head_dim
         self.head_count_kv = head_count_kv
-        self.use_grok = use_grok
+        self.attention_scale = attention_scale
+        self.softcap = softcap
 
         self.add_module(
             "attn_norm", RMSNormLayer(theta("attn_norm"), epsilon=rms_epsilon)
@@ -56,7 +58,12 @@ def __init__(
         self.add_module("attn_v", LinearLayer(theta("attn_v")))
         self.add_module("attn_output", LinearLayer(theta("attn_output")))
 
-        if self.use_grok:
+        if theta.optional_tensor("attn_output_norm") is None:
+            self.add_module(
+                "attn_output_norm",
+                torch.nn.Identity(),
+            )
+        else:
             self.add_module(
                 "attn_output_norm",
                 RMSNormLayer(theta("attn_output_norm"), epsilon=rms_epsilon),
@@ -147,16 +154,16 @@ def repeat_kv(x: torch.Tensor) -> torch.Tensor:
         keys = xk.transpose(1, 2)
         values = xv.transpose(1, 2)
 
+        attn_weights = ops.matmul(xq, keys.transpose(2, 3))
+        if self.attention_scale is None:
+            attn_weights = attn_weights / math.sqrt(self.head_dim)
+        else:
+            attn_weights = attn_weights * self.attention_scale
+
         # Flash attention.
-        if not self.use_grok:
-            attn_weights = ops.matmul(xq, keys.transpose(2, 3)) / math.sqrt(
-                self.head_dim
-            )
-        elif self.use_grok:
-            attn_weights = ops.matmul(xq, keys.transpose(2, 3))
-            attn_weights = 30.0 * torch.tanh(
-                attn_weights * (0.08838834764831845 / 30.0)
-            )
+        if self.softcap is not None:
+            attn_weights = self.softcap * torch.tanh(attn_weights / self.softcap)
+
         self.assert_not_nan(attn_weights)
 
         # Apply attention mask.
@@ -172,12 +179,9 @@ def repeat_kv(x: torch.Tensor) -> torch.Tensor:
 
         # Project.
         attn_output = self.attn_output(attn_output)
-
-        if self.use_grok:
-            attn_output = self.attn_output_norm(attn_output)
+        attn_output = self.attn_output_norm(attn_output)
 
         h = h + attn_output
-
         return h
 
     def transact_cache_direct(