[torch.compile] Enable attention and allreduce fusion without custom ops enabled

.buildkite/test-pipeline.yaml

@@ -416,15 +416,16 @@ steps:
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s compile/piecewise/
 
-- label: PyTorch Fullgraph Test # 20min
-  timeout_in_minutes: 30
+- label: PyTorch Fullgraph Test # 22min
+  timeout_in_minutes: 35
   mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
   - tests/compile
   commands:
   - pytest -v -s compile/test_full_graph.py
+  - pytest -v -s compile/test_fusions_e2e.py
 
 - label: Kernels Core Operation Test # 48min
   timeout_in_minutes: 75
@@ -807,8 +808,8 @@ steps:
     # Whisper needs spawn method to avoid deadlock
     - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
 
-- label: Blackwell Test # 38 min
-  timeout_in_minutes: 60
+- label: Blackwell Test # 21 min
+  timeout_in_minutes: 30
   working_dir: "/vllm-workspace/"
   gpu: b200
   # optional: true
@@ -821,8 +822,6 @@ steps:
   - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
   - vllm/v1/attention/backends/flashinfer.py
-  - vllm/compilation/fusion.py
-  - vllm/compilation/fusion_attn.py
   commands:
     - nvidia-smi
     - python3 examples/offline_inference/basic/chat.py
@@ -839,15 +838,32 @@ steps:
     - pytest -v -s tests/kernels/quantization/test_nvfp4_scaled_mm.py
     - pytest -v -s tests/kernels/quantization/test_flashinfer_scaled_mm.py
     - pytest -v -s tests/kernels/quantization/test_flashinfer_nvfp4_scaled_mm.py
+    - pytest -v -s tests/kernels/quantization/test_nvfp4_qutlass.py
+    - pytest -v -s tests/kernels/quantization/test_mxfp4_qutlass.py
     - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
     - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
-    # Fusion
-    - pytest -v -s tests/compile/test_fusion_all_reduce.py
-    - pytest -v -s tests/compile/test_fusion_attn.py::test_attention_quant_pattern
     - pytest -v -s tests/kernels/moe/test_flashinfer.py
+
+- label: Blackwell Fusion Tests # 30 min
+  timeout_in_minutes: 40
+  working_dir: "/vllm-workspace/"
+  gpu: b200
+  source_file_dependencies:
+  - csrc/quantization/fp4/
+  - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
+  - vllm/v1/attention/backends/flashinfer.py
+  - vllm/compilation/
+  # can affect pattern matching
+  - vllm/model_executor/layers/layernorm.py
+  - vllm/model_executor/layers/activation.py
+  - vllm/model_executor/layers/quantization/input_quant_fp8.py
+  commands:
+    - nvidia-smi
+    - pytest -v -s tests/compile/test_fusion_attn.py
     - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
-    - pytest -v -s tests/kernels/quantization/test_nvfp4_qutlass.py
-    - pytest -v -s tests/kernels/quantization/test_mxfp4_qutlass.py
+    # this runner has 2 GPUs available even though num_gpus=2 is not set
+    - pytest -v -s tests/compile/test_fusion_all_reduce.py
+    - pytest -v -s tests/compile/test_fusions_e2e.py
 
 - label: Blackwell GPT-OSS Eval
   timeout_in_minutes: 60
@@ -1100,14 +1116,16 @@ steps:
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
 
 ##### H200 test #####
-- label: Distrubted Tests (H200) # optional
+- label: Distributed Tests (H200) # optional
   gpu: h200
   optional: true
   working_dir: "/vllm-workspace/"
   num_gpus: 2
   commands:
     - pytest -v -s tests/compile/test_async_tp.py
     - pytest -v -s tests/compile/test_sequence_parallelism.py
+    - pytest -v -s tests/compile/test_fusion_all_reduce.py
+    - pytest -v -s tests/compile/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
     - pytest -v -s tests/distributed/test_context_parallel.py
     - CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1  --dp-size=2 --max-model-len 2048
 

csrc/layernorm_kernels.cu

@@ -392,6 +392,8 @@ void fused_add_rms_norm(torch::Tensor& input,     // [..., hidden_size]
                         torch::Tensor& residual,  // [..., hidden_size]
                         torch::Tensor& weight,    // [hidden_size]
                         double epsilon) {
+  TORCH_CHECK(weight.scalar_type() == input.scalar_type());
+  TORCH_CHECK(input.scalar_type() == residual.scalar_type());
   TORCH_CHECK(residual.is_contiguous());
   TORCH_CHECK(weight.is_contiguous());
   int hidden_size = input.size(-1);

csrc/layernorm_quant_kernels.cu

@@ -229,6 +229,8 @@ void fused_add_rms_norm_static_fp8_quant(
     double epsilon) {
   TORCH_CHECK(out.is_contiguous());
   TORCH_CHECK(residual.is_contiguous());
+  TORCH_CHECK(residual.scalar_type() == input.scalar_type());
+  TORCH_CHECK(weight.scalar_type() == input.scalar_type());
   int hidden_size = input.size(-1);
   int input_stride = input.stride(-2);
   int num_tokens = input.numel() / hidden_size;

csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu

@@ -145,7 +145,11 @@ void rms_norm_dynamic_per_token_quant(
   if (scale_ub.has_value()) {
     TORCH_CHECK(out.dtype() == kFp8Type);
   }
+  TORCH_CHECK(weight.dtype() == input.dtype());
   TORCH_CHECK(scales.dtype() == torch::kFloat32);
+  if (residual) {
+    TORCH_CHECK(residual->scalar_type() == input.scalar_type());
+  }
 
   VLLM_DISPATCH_FLOATING_TYPES(
       input.scalar_type(), "rms_norm_dynamic_per_token_quant_dispatch", [&] {

tests/compile/backend.py

@@ -3,16 +3,22 @@
 
 import weakref
 from collections.abc import Callable, Sequence
+from contextlib import nullcontext
 from copy import deepcopy
 
+import depyf
 from torch import fx
 from torch._ops import OpOverload
+from torch.fx._utils import lazy_format_graph_code
 
 from vllm.compilation.fx_utils import find_op_nodes
 from vllm.compilation.inductor_pass import InductorPass
 from vllm.compilation.pass_manager import with_pattern_match_debug
 from vllm.compilation.vllm_inductor_pass import VllmInductorPass
 from vllm.config import VllmConfig, get_current_vllm_config
+from vllm.logger import init_logger
+
+logger = init_logger("vllm.tests.compile.backend")
 
 
 class LazyInitPass(InductorPass):
@@ -45,20 +51,32 @@ class TestBackend:
 
     def __init__(self, *passes: InductorPass | Callable[[fx.Graph], None]):
         self.custom_passes = list(passes)
-        compile_config = get_current_vllm_config().compilation_config
-        self.inductor_config = compile_config.inductor_compile_config
+        vllm_config = get_current_vllm_config()
+        compile_config = vllm_config.compilation_config
+        # Deepcopy to allow multiple TestBackend instances to use the same VllmConfig
+        self.inductor_config = deepcopy(compile_config.inductor_compile_config)
         self.inductor_config["force_disable_caches"] = True
         self.inductor_config["post_grad_custom_post_pass"] = self.post_pass
 
+        if debug_dump_path := vllm_config.compile_debug_dump_path():
+            logger.debug("Dumping depyf output to %s", debug_dump_path)
+            self.debug_ctx = depyf.prepare_debug(debug_dump_path.as_posix())
+        else:
+            self.debug_ctx = nullcontext()
+
     def __call__(self, graph: fx.GraphModule, example_inputs):
         self.graph_pre_compile = deepcopy(graph)
         from torch._inductor.compile_fx import compile_fx
 
-        return compile_fx(graph, example_inputs, config_patches=self.inductor_config)
+        with self.debug_ctx:
+            return compile_fx(
+                graph, example_inputs, config_patches=self.inductor_config
+            )
 
     @with_pattern_match_debug
     def post_pass(self, graph: fx.Graph):
         self.graph_pre_pass = deepcopy(graph)
+        lazy_format_graph_code("graph_pre_pass", graph.owning_module)
 
         VllmInductorPass.dump_prefix = 0
         for pass_ in self.custom_passes:
@@ -68,6 +86,7 @@ def post_pass(self, graph: fx.Graph):
         VllmInductorPass.dump_prefix = None
 
         self.graph_post_pass = deepcopy(graph)
+        lazy_format_graph_code("graph_post_pass", graph.owning_module)
         # assign by reference, will reflect the final state of the graph
         self.final_graph = graph