[https://nvbugs/5550409][fix] Disable torch compile in piecewise attention part to Avoid host overhead (NVIDIA#8708)

Signed-off-by: yizhang-nv <[email protected]>

Author: yizhang-nv

tensorrt_llm/_torch/compilation/piecewise_optimizer.py

@@ -13,7 +13,8 @@
 
 from ..utils import (get_model_extra_attrs,
                      get_per_request_piecewise_cuda_graph_flag,
-                     get_piecewise_cuda_graph_flag, make_weak_ref)
+                     get_piecewise_cuda_graph_flag, make_weak_ref,
+                     set_piecewise_running)
 from .multi_stream.auto_multi_stream import multi_stream_schedule
 from .utils import get_capture_piecewise_cuda_graph_flag, is_call_function
 
@@ -27,6 +28,7 @@ def __init__(
         compile_time_num_tokens: Union[int | torch.SymInt],
         capture_num_tokens: list[int],
         exclude_modules_id: list[int],
+        piecewise_runner_num: int,
         graph_pool_handle: tuple[int, int],
         garbage_collect_values: bool = True,
         graph=None,
@@ -38,6 +40,8 @@ def __init__(
 
         self.compile_time_num_tokens = compile_time_num_tokens
         self.capture_num_tokens = capture_num_tokens
+        self.piecewise_runner_num = piecewise_runner_num
+        self.piecewise_runner_idx = 0
         self.exclude_modules = [f"submod_{i}" for i in exclude_modules_id]
         self.graph_pool_handle = graph_pool_handle
         self.enable_inductor = enable_inductor
@@ -90,8 +94,10 @@ def call_module(self, target, args, kwargs):
                 self.graph_pool_handle,
                 compile_fx(submod, args) if self.enable_inductor else submod,
                 self.enable_inductor,
+                self.piecewise_runner_idx == 0,
+                self.piecewise_runner_idx == self.piecewise_runner_num - 1,
             )
-
+            self.piecewise_runner_idx += 1
         return output
 
 
@@ -124,6 +130,8 @@ def __init__(
         graph_pool_handle,
         default_callable: Callable,
         enable_inductor: bool,
+        is_first_runner: bool,
+        is_last_runner: bool,
     ):
         if runtime_num_tokens_idx != None:
             assert isinstance(compile_time_num_tokens, torch.SymInt)
@@ -138,6 +146,8 @@ def __init__(
         self.enable_inductor = enable_inductor
 
         self.entries: dict[int, Entry] = {}
+        self.is_first_runner = is_first_runner
+        self.is_last_runner = is_last_runner
 
         for num_tokens in capture_num_tokens:
             self.entries[num_tokens] = Entry(
@@ -161,6 +171,12 @@ def __call__(self, *args):
                 or not get_per_request_piecewise_cuda_graph_flag()):
             return self.default_callable(*args)
 
+        if self.is_first_runner or self.is_last_runner:
+            if self.is_first_runner == self.is_last_runner:
+                set_piecewise_running(False)
+            else:
+                set_piecewise_running(self.is_first_runner)
+
         entry = self.entries[runtime_num_of_token]
 
         if entry.enable_inductor and not entry.compiled:
@@ -267,6 +283,7 @@ def piecewise_optimizer(
         input_num_tokens,
         capture_num_tokens,
         exclude_modules_id,
+        len(set(node_to_graph_id.values())) - len(exclude_modules_id),
         graph_pool_handle,
         max_num_streams=max_num_streams,
     )

tensorrt_llm/_torch/modules/attention.py

@@ -23,7 +23,7 @@
 from ..model_config import ModelConfig
 from ..peft.lora.layer import LoraLayer, LoraModuleType
 from ..utils import (Fp4QuantizedTensor, get_model_extra_attrs,
-                     is_torch_compiling)
+                     is_piecewise_running, is_torch_compiling)
 from .linear import Linear, TensorParallelMode, WeightMode, WeightsLoadingConfig
 from .multi_stream_utils import maybe_execute_in_parallel
 from .rms_norm import RMSNorm
@@ -76,13 +76,24 @@ def extract_extra_attrs(layer_idx: str, attn_type: str):
     return metadata, attn_layer
 
 
-@torch.compile
-def compiled_copy_(dst, src):
+def maybe_compile(func):
+
+    def wrapper(*args, **kwargs):
+        if is_piecewise_running():
+            # When piecewise running, we don't need to compile the function to avoid host overhead in attention op.
+            return func(*args, **kwargs)
+        return torch.compile(func)(*args, **kwargs)
+
+    return wrapper
+
+
+@maybe_compile
+def maybe_compiled_copy_(dst, src):
     dst.copy_(src)
 
 
-@torch.compile
-def compiled_cat(tensors, dim):
+@maybe_compile
+def maybe_compiled_cat(tensors, dim):
     return torch.cat(tensors, dim)
 
 
@@ -1222,8 +1233,9 @@ def forward_context_default(
         )
 
         k = torch.empty_like(q).view(-1, self.num_heads, self.qk_head_dim)
-        compiled_copy_(k[..., :self.qk_nope_head_dim],
-                       k_nope.view(-1, self.num_heads, self.qk_nope_head_dim))
+        maybe_compiled_copy_(
+            k[..., :self.qk_nope_head_dim],
+            k_nope.view(-1, self.num_heads, self.qk_nope_head_dim))
         if self.apply_rotary_emb:
             k[..., self.qk_nope_head_dim:] = k_pe.view(-1, 1,
                                                        self.qk_rope_head_dim)
@@ -1317,7 +1329,7 @@ def forward_context_with_cached_kv(
         full_k_nope = full_k_nope.view(-1, self.num_heads,
                                        self.qk_nope_head_dim)
         full_k_pe = full_k_pe.view(-1, 1, self.qk_rope_head_dim)
-        full_k = compiled_cat(
+        full_k = maybe_compiled_cat(
             (full_k_nope, full_k_pe.expand(-1, self.num_heads, -1)), dim=-1)
         full_k = full_k.view(-1, self.num_heads * self.qk_head_dim)
 
@@ -1412,7 +1424,7 @@ def forward_context_with_chunked_prefill(
             chunked_k_nope = chunked_k_nope.view(-1, self.num_heads,
                                                  self.qk_nope_head_dim)
             chunked_k_pe = chunked_k_pe.view(-1, 1, self.qk_rope_head_dim)
-            chunked_k = compiled_cat(
+            chunked_k = maybe_compiled_cat(
                 (chunked_k_nope, chunked_k_pe.expand(-1, self.num_heads, -1)),
                 dim=-1)
             chunked_k = chunked_k.view(-1, self.num_heads * self.qk_head_dim)
@@ -1470,7 +1482,8 @@ def forward_context_with_chunked_prefill(
 
         k_nope = k_nope.view(-1, self.num_heads, self.qk_nope_head_dim)
         k_pe = k_pe.view(-1, 1, self.qk_rope_head_dim)
-        k = compiled_cat((k_nope, k_pe.expand(-1, self.num_heads, -1)), dim=-1)
+        k = maybe_compiled_cat((k_nope, k_pe.expand(-1, self.num_heads, -1)),
+                               dim=-1)
         k = k.view(-1, self.num_heads * self.qk_head_dim)
 
         # copy q_lens to replace kv_lens_runtime

tensorrt_llm/_torch/utils.py

@@ -12,6 +12,7 @@
 from tensorrt_llm.quantization.utils import fp4_utils
 
 is_torch_compiling_flag = False
+is_piecewise_running_flag = False
 
 aux_stream_name_list = [
     'Attention',
@@ -40,6 +41,16 @@ def is_torch_compiling() -> bool:
     return is_torch_compiling_flag
 
 
+def set_piecewise_running(enable: bool):
+    global is_piecewise_running_flag
+    is_piecewise_running_flag = enable
+
+
+def is_piecewise_running() -> bool:
+    global is_piecewise_running_flag
+    return is_piecewise_running_flag
+
+
 _global_attrs = threading.local()