[Hardware][TPU] Support MoE with Pallas GMM kernel (vllm-project#6457)

Author: WoosukKwon

Dockerfile.tpu

@@ -1,11 +1,13 @@
-ARG NIGHTLY_DATE="20240601"
+ARG NIGHTLY_DATE="20240713"
 ARG BASE_IMAGE="us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.10_tpuvm_$NIGHTLY_DATE"
 
 FROM $BASE_IMAGE
 WORKDIR /workspace
 
 # Install aiohttp separately to avoid build errors.
 RUN pip install aiohttp
+# Install NumPy 1 instead of NumPy 2.
+RUN pip install "numpy<2"
 # Install the TPU and Pallas dependencies.
 RUN pip install torch_xla[tpu] -f https://storage.googleapis.com/libtpu-releases/index.html
 RUN pip install torch_xla[pallas] -f https://storage.googleapis.com/jax-releases/jax_nightly_releases.html -f https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html

docs/source/getting_started/tpu-installation.rst

@@ -56,7 +56,7 @@ First, install the dependencies:
     $ pip uninstall torch torch-xla -y
 
     $ # Install PyTorch and PyTorch XLA.
-    $ export DATE="+20240601"
+    $ export DATE="+20240713"
     $ pip install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-nightly${DATE}-cp310-cp310-linux_x86_64.whl
     $ pip install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-nightly${DATE}-cp310-cp310-linux_x86_64.whl
 
@@ -85,7 +85,7 @@ Next, build vLLM from source. This will only take a few seconds:
         ImportError: libopenblas.so.0: cannot open shared object file: No such file or directory
 
 
-    You can install OpenBLAS with the following command:
+    Please install OpenBLAS with the following command:
 
     .. code-block:: console
 

vllm/model_executor/layers/fused_moe/layer.py

@@ -104,6 +104,24 @@ def forward_cpu(self, *args, **kwargs):
         raise NotImplementedError(
             "The CPU backend currently does not support MoE.")
 
+    def forward_tpu(
+        self,
+        x: torch.Tensor,
+        w1: torch.Tensor,
+        w2: torch.Tensor,
+        router_logits: torch.Tensor,
+        top_k: int,
+        renormalize: bool,
+        use_grouped_topk: bool,
+        num_expert_group: Optional[int],
+        topk_group: Optional[int],
+    ) -> torch.Tensor:
+        from vllm.model_executor.layers.fused_moe.moe_pallas import fused_moe
+        assert not use_grouped_topk
+        assert num_expert_group is None
+        assert topk_group is None
+        return fused_moe(x, w1, w2, router_logits, top_k, renormalize)
+
 
 class FusedMoE(torch.nn.Module):
     """FusedMoE layer for MoE models.

vllm/model_executor/layers/fused_moe/moe_pallas.py

@@ -0,0 +1,62 @@
+import torch
+import torch.nn.functional as F
+from torch_xla.experimental.custom_kernel import _histogram
+
+
+def fused_moe(
+    hidden_states: torch.Tensor,
+    w1: torch.Tensor,
+    w2: torch.Tensor,
+    gating_output: torch.Tensor,
+    topk: int,
+    renormalize: bool,
+) -> torch.Tensor:
+    """
+    Args:
+        hidden_states: [*, hidden_size]
+        w1: [num_experts, intermediate_size * 2, hidden_size]
+        w2: [num_experts, hidden_size, intermediate_size]
+        gating_output: [*, num_experts]
+    """
+    orig_shape = hidden_states.shape
+    hidden_size = hidden_states.shape[-1]
+    num_tokens = hidden_states.shape[:-1].numel()
+    num_experts = w1.shape[0]
+    intermediate_size = w2.shape[-1]
+    device = hidden_states.device
+    dtype = hidden_states.dtype
+    assert (num_tokens * topk) % 16 == 0, (
+        "The Pallas GMM kernel requires num_tokens * topk to be a multiple of "
+        f"16 but got {num_tokens * topk}")
+
+    hidden_states = hidden_states.view(num_tokens, hidden_size)
+    gating_output = gating_output.view(num_tokens, num_experts)
+    topk_weights = gating_output.softmax(dim=-1, dtype=torch.float)
+    topk_weights, topk_indices = topk_weights.topk(topk, dim=-1)
+    if renormalize:
+        topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
+    topk_weights = topk_weights.to(dtype)
+
+    topk_indices = topk_indices.flatten()
+    topk_argsort_indices = topk_indices.argsort()
+    topk_argsort_revert_indices = topk_argsort_indices.argsort()
+    token_indices = torch.arange(num_tokens,
+                                 device=device).repeat_interleave(topk)
+    token_indices = token_indices[topk_argsort_indices]
+    group_sizes = _histogram(topk_indices.to(torch.int32), 0, num_experts - 1)
+
+    # NOTE(woosuk): The GMM Pallas kernel requires a different weight layout
+    # from HF Transformers.
+    w1 = w1.transpose(1, 2)
+    w2 = w2.transpose(1, 2)
+
+    x = hidden_states[token_indices]
+    x = torch.ops.xla.gmm(x, w1, group_sizes)
+    x = F.silu(x[..., :intermediate_size]) * x[..., intermediate_size:]
+    x = torch.ops.xla.gmm(x, w2, group_sizes)
+    x = x[topk_argsort_revert_indices].reshape(-1, topk, hidden_size)
+
+    x = x * topk_weights.unsqueeze_(dim=-1)
+    x = x.sum(dim=-2)
+    x = x.reshape(orig_shape)
+    return x

vllm/worker/tpu_model_runner.py

@@ -598,11 +598,10 @@ def _get_padded_prefill_len(x: int) -> int:
 
 
 def _get_padded_batch_size(batch_size: int) -> int:
-    if batch_size <= 2:
-        return batch_size
-    elif batch_size <= 4:
-        return 4
-    elif batch_size <= 8:
+    # The GMM Pallas kernel requires num_tokens * topk to be a multiple of 16.
+    # To meet this requirement in the simplest way, we set the minimal batch
+    # size to 8.
+    if batch_size <= 8:
         return 8
     else:
         return ((batch_size + 15) // 16) * 16