sglang 0.2.11__py3-none-any.whl → 0.2.13__py3-none-any.whl

sglang/srt/layers/activation.py

@@ -0,0 +1,32 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+    http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
+"""Fused operators for activation layers."""
+
+import torch
+import torch.nn.functional as F
+from flashinfer.activation import silu_and_mul
+from vllm.model_executor.custom_op import CustomOp
+
+
+class SiluAndMul(CustomOp):
+    def forward_native(self, x: torch.Tensor) -> torch.Tensor:
+        d = x.shape[-1] // 2
+        return F.silu(x[..., :d]) * x[..., d:]
+
+    def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        silu_and_mul(x, out)
+        return out

sglang/srt/layers/{token_attention.py → decode_attention.py}

@@ -13,6 +13,10 @@ See the License for the specific language governing permissions and
 limitations under the License.
 """
 
+"""
+Memory-efficient attention for decoding.
+"""
+
 # Adapted from
 # https://github.com/ModelTC/lightllm/blob/f2a54f0912293f683bf1d1695fd12c4098a5bf82/lightllm/models/llama/triton_kernel/token_attention_nopad_att1.py
 # https://github.com/ModelTC/lightllm/blob/f2a54f0912293f683bf1d1695fd12c4098a5bf82/lightllm/models/llama/triton_kernel/token_attention_softmax_and_reducev.py
@@ -194,7 +198,7 @@ def _fwd_kernel_stage2(
     tl.store(out_ptrs, acc)
 
 
-def _token_att_m_fwd(
+def _decode_att_m_fwd(
     q,
     k_buffer,
     att_out,
@@ -254,7 +258,7 @@ def _token_att_m_fwd(
     )
 
 
-def _token_softmax_reducev_fwd(
+def _decode_softmax_reducev_fwd(
     logics,
     v_buffer,
     o,
@@ -292,7 +296,7 @@ def _token_softmax_reducev_fwd(
     )
 
 
-def token_attention_fwd(
+def decode_attention_fwd(
     q,
     k_buffer,
     v_buffer,
@@ -312,7 +316,7 @@ def token_attention_fwd(
             (q.shape[-2], total_num_tokens), dtype=REDUCE_TORCH_TYPE, device="cuda"
         )
 
-    _token_att_m_fwd(
+    _decode_att_m_fwd(
         q,
         k_buffer,
         att_m,
@@ -324,7 +328,7 @@ def token_attention_fwd(
         sm_scale,
         logit_cap,
     )
-    _token_softmax_reducev_fwd(
+    _decode_softmax_reducev_fwd(
         att_m,
         v_buffer,
         o,

sglang/srt/layers/extend_attention.py

@@ -13,11 +13,16 @@ See the License for the specific language governing permissions and
 limitations under the License.
 """
 
+"""
+Memory-efficient attention for prefill.
+It supporst page size = 1 and prefill with KV cache (i.e. extend).
+"""
+
 import torch
 import triton
 import triton.language as tl
 
-from sglang.srt.layers.context_flashattention_nopad import context_attention_fwd
+from sglang.srt.layers.prefill_attention import context_attention_fwd
 
 CUDA_CAPABILITY = torch.cuda.get_device_capability()
 
@@ -270,7 +275,9 @@ def extend_attention_fwd(
         BLOCK_DPE = 0
     BLOCK_DV = Lv
 
-    if CUDA_CAPABILITY[0] >= 8:
+    if CUDA_CAPABILITY[0] >= 9:
+        BLOCK_M, BLOCK_N = (128, 64)
+    elif CUDA_CAPABILITY[0] >= 8:
         BLOCK_M, BLOCK_N = (128, 128) if Lq <= 128 else (64, 64)
     else:
         BLOCK_M, BLOCK_N = (64, 64) if Lq <= 128 else (32, 32)

sglang/srt/layers/fused_moe/__init__.py

@@ -0,0 +1 @@
+from sglang.srt.layers.fused_moe.layer import FusedMoE, FusedMoEMethodBase

sglang/srt/layers/{fused_moe.py → fused_moe/fused_moe.py}

@@ -1,20 +1,5 @@
-"""
-Copyright 2023-2024 SGLang Team
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-"""
-
 # Adapted from
-# https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/layers/fused_moe/fused_moe.py#L1
+# https://github.com/vllm-project/vllm/tree/v0.5.4/vllm/model_executor/layers/fused_moe
 """Fused MoE kernel."""
 import functools
 import json
@@ -24,6 +9,7 @@ from typing import Any, Dict, Optional, Tuple
 import torch
 import triton
 import triton.language as tl
+import vllm.envs as envs
 from vllm import _custom_ops as ops
 from vllm.logger import init_logger
 
@@ -373,6 +359,31 @@ def get_default_config(
     return config
 
 
+def try_get_optimal_moe_config(
+    w1_shape: Tuple[int, ...],
+    w2_shape: Tuple[int, ...],
+    top_k: int,
+    dtype: Optional[str],
+    M: int,
+    override_config: Optional[Dict[str, Any]] = None,
+):
+    if override_config:
+        config = override_config
+    else:
+        # First try to load optimal config from the file
+        E, _, N = w2_shape
+        configs = get_moe_configs(E, N, dtype)
+
+        if configs:
+            # If an optimal configuration map has been found, look up the
+            # optimal config
+            config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
+        else:
+            # Else use the default config
+            config = get_default_config(M, E, N, w1_shape[2], top_k, dtype)
+    return config
+
+
 def fused_topk(
     hidden_states: torch.Tensor,
     gating_output: torch.Tensor,
@@ -403,6 +414,41 @@ def fused_topk(
     return topk_weights, topk_ids
 
 
+# This is used by the Deepseek-V2 model
+def grouped_topk(
+    hidden_states: torch.Tensor,
+    gating_output: torch.Tensor,
+    topk: int,
+    renormalize: bool,
+    num_expert_group: int = 0,
+    topk_group: int = 0,
+):
+
+    assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
+
+    scores = torch.softmax(gating_output, dim=-1)
+    num_token = scores.shape[0]
+    group_scores = (
+        scores.view(num_token, num_expert_group, -1).max(dim=-1).values
+    )  # [n, n_group]
+    group_idx = torch.topk(group_scores, k=topk_group, dim=-1, sorted=False)[
+        1
+    ]  # [n, top_k_group]
+    group_mask = torch.zeros_like(group_scores)  # [n, n_group]
+    group_mask.scatter_(1, group_idx, 1)  # [n, n_group]
+    score_mask = (
+        group_mask.unsqueeze(-1)
+        .expand(num_token, num_expert_group, scores.shape[-1] // num_expert_group)
+        .reshape(num_token, -1)
+    )  # [n, e]
+    tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0)  # [n, e]
+    topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)
+
+    if renormalize:
+        topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
+    return topk_weights, topk_ids
+
+
 def fused_experts(
     hidden_states: torch.Tensor,
     w1: torch.Tensor,
@@ -425,24 +471,23 @@ def fused_experts(
     assert w2.is_contiguous(), "Expert weights2 must be contiguous"
     assert hidden_states.dtype in [torch.float32, torch.float16, torch.bfloat16]
 
-    M, _ = hidden_states.shape
+    num_tokens, _ = hidden_states.shape
     E, N, _ = w1.shape
+    # We execute the fused_moe kernel in chunks to circumvent this issue:
+    # https://github.com/vllm-project/vllm/issues/5938
+    CHUNK_SIZE = envs.VLLM_FUSED_MOE_CHUNK_SIZE
+    M = min(num_tokens, CHUNK_SIZE)
+
+    get_config_func = functools.partial(
+        try_get_optimal_moe_config,
+        w1.shape,
+        w2.shape,
+        topk_ids.shape[1],
+        "float8" if use_fp8 else None,
+        override_config=override_config,
+    )
 
-    if override_config:
-        config = override_config
-    else:
-        # First try to load optimal config from the file
-        configs = get_moe_configs(E, w2.shape[2], "float8" if use_fp8 else None)
-
-        if configs:
-            # If an optimal configuration map has been found, look up the
-            # optimal config
-            config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
-        else:
-            # Else use the default config
-            config = get_default_config(
-                M, E, N, w1.shape[2], topk_ids.shape[1], "float8" if use_fp8 else None
-            )
+    config = get_config_func(M)
 
     intermediate_cache1 = torch.empty(
         (M, topk_ids.shape[1], N),
@@ -460,56 +505,85 @@ def fused_experts(
         dtype=hidden_states.dtype,
     )
 
-    sorted_token_ids, expert_ids, num_tokens_post_padded = moe_align_block_size(
-        topk_ids, config["BLOCK_SIZE_M"], E
-    )
     compute_type = tl.bfloat16 if hidden_states.dtype == torch.bfloat16 else tl.float16
 
-    invoke_fused_moe_kernel(
-        hidden_states,
-        w1,
-        intermediate_cache1,
-        a1_scale,
-        w1_scale,
-        topk_weights,
-        topk_ids,
-        sorted_token_ids,
-        expert_ids,
-        num_tokens_post_padded,
-        False,
-        topk_ids.shape[1],
-        config,
-        compute_type=compute_type,
-        use_fp8=use_fp8,
-    )
+    if inplace:
+        out_hidden_states = hidden_states
+    else:
+        out_hidden_states = torch.empty_like(hidden_states)
 
-    ops.gelu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
+    for chunk in range((num_tokens // CHUNK_SIZE) + 1):
+        begin_chunk_idx, end_chunk_idx = (
+            chunk * CHUNK_SIZE,
+            min((chunk + 1) * CHUNK_SIZE, num_tokens),
+        )
+        curr_hidden_states = hidden_states[begin_chunk_idx:end_chunk_idx]
+        tokens_in_chunk, _ = curr_hidden_states.shape
+
+        if tokens_in_chunk == 0:
+            break
+
+        if tokens_in_chunk < CHUNK_SIZE and chunk > 0:
+            # Adjust the intermediate cache size and config for the last
+            # chunk. Note that in most cases we only have one chunk
+            # so the cache size and config are already set correctly and
+            # do not need to be adjusted.
+            intermediate_cache1 = intermediate_cache1[:tokens_in_chunk]
+            intermediate_cache2 = intermediate_cache2[:tokens_in_chunk]
+            intermediate_cache3 = intermediate_cache3[:tokens_in_chunk]
+            config = get_config_func(tokens_in_chunk)
+
+        curr_topk_ids = topk_ids[begin_chunk_idx:end_chunk_idx]
+        curr_topk_weights = topk_weights[begin_chunk_idx:end_chunk_idx]
+
+        sorted_token_ids, expert_ids, num_tokens_post_padded = moe_align_block_size(
+            curr_topk_ids, config["BLOCK_SIZE_M"], E
+        )
 
-    invoke_fused_moe_kernel(
-        intermediate_cache2,
-        w2,
-        intermediate_cache3,
-        a2_scale,
-        w2_scale,
-        topk_weights,
-        topk_ids,
-        sorted_token_ids,
-        expert_ids,
-        num_tokens_post_padded,
-        True,
-        1,
-        config,
-        compute_type=compute_type,
-        use_fp8=use_fp8,
-    )
+        invoke_fused_moe_kernel(
+            curr_hidden_states,
+            w1,
+            intermediate_cache1,
+            a1_scale,
+            w1_scale,
+            curr_topk_weights,
+            curr_topk_ids,
+            sorted_token_ids,
+            expert_ids,
+            num_tokens_post_padded,
+            False,
+            topk_ids.shape[1],
+            config,
+            compute_type=compute_type,
+            use_fp8=use_fp8,
+        )
 
-    if inplace:
-        return torch.sum(
+        ops.gelu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
+
+        invoke_fused_moe_kernel(
+            intermediate_cache2,
+            w2,
+            intermediate_cache3,
+            a2_scale,
+            w2_scale,
+            curr_topk_weights,
+            curr_topk_ids,
+            sorted_token_ids,
+            expert_ids,
+            num_tokens_post_padded,
+            True,
+            1,
+            config,
+            compute_type=compute_type,
+            use_fp8=use_fp8,
+        )
+
+        torch.sum(
             intermediate_cache3.view(*intermediate_cache3.shape),
             dim=1,
-            out=hidden_states,
+            out=out_hidden_states[begin_chunk_idx:end_chunk_idx],
         )
-    return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
+    return out_hidden_states
 
 
 def fused_moe(
@@ -521,6 +595,9 @@ def fused_moe(
     renormalize: bool,
     inplace: bool = False,
     override_config: Optional[Dict[str, Any]] = None,
+    use_grouped_topk: bool = False,
+    num_expert_group: Optional[int] = None,
+    topk_group: Optional[int] = None,
     use_fp8: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
@@ -543,6 +620,10 @@ def fused_moe(
         Defaults to False.
     - override_config (Optional[Dict[str, Any]]): Optional override
         for the kernel configuration.
+    - num_expert_group: Optional[int]: additional parameter for grouped_topk
+    - topk_group: Optional[int]: additional parameter for grouped_topk
+    - use_grouped_topk: If True, use grouped_topk instead of fused_topk
+        note: Deepseekv2 model uses grouped_topk
     - use_fp8 (bool): If True, use fp8 arithmetic to compute the inner
         products for w1 and w2. Defaults to False.
     - w1_scale (Optional[torch.Tensor]): Optional scale to be used for
@@ -556,12 +637,18 @@ def fused_moe(
     # Check constraints.
     assert gating_output.shape[1] == w1.shape[0], "Number of experts mismatch"
 
-    if hasattr(ops, "topk_softmax"):
-        topk_weights, topk_ids = fused_topk(
-            hidden_states, gating_output, topk, renormalize
+    if use_grouped_topk:
+        assert num_expert_group is not None and topk_group is not None
+        topk_weights, topk_ids = grouped_topk(
+            hidden_states,
+            gating_output,
+            topk,
+            renormalize,
+            num_expert_group,
+            topk_group,
         )
     else:
-        topk_weights, topk_ids = fused_topk_v0_4_3(
+        topk_weights, topk_ids = fused_topk(
             hidden_states, gating_output, topk, renormalize
         )
 
@@ -579,33 +666,3 @@ def fused_moe(
         a1_scale=a1_scale,
         a2_scale=a2_scale,
     )
-
-
-def fused_topk_v0_4_3(
-    hidden_states: torch.Tensor,
-    gating_output: torch.Tensor,
-    topk: int,
-    renormalize: bool,
-):
-    import vllm._moe_C as moe_kernels
-
-    M, _ = hidden_states.shape
-
-    topk_weights = torch.empty(
-        M, topk, dtype=torch.float32, device=hidden_states.device
-    )
-    topk_ids = torch.empty(M, topk, dtype=torch.int32, device=hidden_states.device)
-    token_expert_indicies = torch.empty(
-        M, topk, dtype=torch.int32, device=hidden_states.device
-    )
-    moe_kernels.topk_softmax(
-        topk_weights,
-        topk_ids,
-        token_expert_indicies,
-        gating_output.float(),  # TODO(woosuk): Optimize this.
-    )
-    del token_expert_indicies  # Not used. Will be used in the future.
-    if renormalize:
-        topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
-
-    return topk_weights, topk_ids