Refactor llm perf backend handling

.github/workflows/update_llm_perf_cpu_pytorch.yaml

@@ -4,16 +4,32 @@ on:
   workflow_dispatch:
   schedule:
     - cron: "0 0 * * *"
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - main
+    types:
+      - opened
+      - reopened
+      - synchronize
+      - labeled
+      - unlabeled
 
 concurrency:
   cancel-in-progress: true
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
 
 env:
   IMAGE: ghcr.io/huggingface/optimum-benchmark:latest-cpu
 
 jobs:
   run_benchmarks:
+    if: ${{
+      (github.event_name == 'push') ||
+      (github.event_name == 'workflow_dispatch') ||
+      contains( github.event.pull_request.labels.*.name, 'leaderboard')}}
     strategy:
       fail-fast: false
       matrix:
@@ -49,4 +65,4 @@ jobs:
             pip install packaging && pip install einops scipy optimum codecarbon
             pip install -U transformers huggingface_hub[hf_transfer]
             pip install -e .
-            python llm_perf/update_llm_perf_cpu_pytorch.py
+            python llm_perf/benchmark_runners/update_llm_perf_cpu_pytorch.py

.github/workflows/update_llm_perf_cuda_pytorch.yaml

@@ -4,16 +4,33 @@ on:
   workflow_dispatch:
   schedule:
     - cron: "0 0 * * *"
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - main
+    types:
+      - opened
+      - reopened
+      - synchronize
+      - labeled
+      - unlabeled
 
 concurrency:
   cancel-in-progress: true
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
 
 env:
   IMAGE: ghcr.io/huggingface/optimum-benchmark:latest-cuda
 
 jobs:
   run_benchmarks:
+    if: ${{
+      (github.event_name == 'push') ||
+      (github.event_name == 'workflow_dispatch') ||
+      contains( github.event.pull_request.labels.*.name, 'leaderboard')}}
+
     strategy:
       fail-fast: false
       matrix:
@@ -54,4 +71,4 @@ jobs:
             pip install packaging && pip install flash-attn einops scipy auto-gptq optimum bitsandbytes autoawq codecarbon
             pip install -U transformers huggingface_hub[hf_transfer]
             pip install -e .
-            python llm_perf/update_llm_perf_cuda_pytorch.py
+            python llm_perf/benchmark_runners/update_llm_perf_cuda_pytorch.py

.github/workflows/update_llm_perf_leaderboard.yaml

@@ -4,13 +4,30 @@ on:
   workflow_dispatch:
   schedule:
     - cron: "0 */6 * * *"
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - main
+    types:
+      - opened
+      - reopened
+      - synchronize
+      - labeled
+      - unlabeled
 
 concurrency:
   cancel-in-progress: true
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
 
 jobs:
   update_llm_perf_leaderboard:
+    if: ${{
+      (github.event_name == 'push') ||
+      (github.event_name == 'workflow_dispatch') ||
+      contains( github.event.pull_request.labels.*.name, 'leaderboard')}}
+
     runs-on: ubuntu-latest
     steps:
       - name: Checkout

llm_perf/benchmark_runners/update_llm_perf_cpu_pytorch.py

@@ -0,0 +1,71 @@
+from typing import Any, Dict, List
+
+from llm_perf.common.benchmark_runner import BenchmarkRunner
+from llm_perf.common.utils import GENERATE_KWARGS, INPUT_SHAPES
+from optimum_benchmark import PyTorchConfig
+from optimum_benchmark.benchmark.config import BenchmarkConfig
+from optimum_benchmark.launchers.process.config import ProcessConfig
+from optimum_benchmark.scenarios.inference.config import InferenceConfig
+
+
+class CPUPyTorchBenchmarkRunner(BenchmarkRunner):
+    def __init__(self):
+        super().__init__(backend="pytorch", hardware="cpu")
+
+    def get_benchmark_config(self, model: str, attn_implementation: str, weights_config: str) -> BenchmarkConfig:
+        assert (
+            weights_config in self.weights_configs
+        ), f"your config does not contain {weights_config}, adjust your _get_weights_configs to fix this issue"
+
+        torch_dtype = self.weights_configs[weights_config]["torch_dtype"]
+        quant_scheme = self.weights_configs[weights_config]["quant_scheme"]
+        quant_config = self.weights_configs[weights_config]["quant_config"]
+
+        launcher_config = ProcessConfig()
+        scenario_config = InferenceConfig(
+            memory=True,
+            energy=True,
+            latency=True,
+            duration=10,
+            iterations=10,
+            warmup_runs=10,
+            input_shapes=INPUT_SHAPES,
+            generate_kwargs=GENERATE_KWARGS,
+        )
+        backend_config = PyTorchConfig(
+            model=model,
+            device="cpu",
+            no_weights=True,
+            library="transformers",
+            task="text-generation",
+            torch_dtype=torch_dtype,
+            quantization_scheme=quant_scheme,
+            quantization_config=quant_config,
+            attn_implementation=attn_implementation,
+            model_kwargs={"trust_remote_code": True},
+        )
+
+        return BenchmarkConfig(
+            name=f"{weights_config}-{attn_implementation}",
+            scenario=scenario_config,
+            launcher=launcher_config,
+            backend=backend_config,
+        )
+
+    def _get_weights_configs(self, subset) -> Dict[str, Dict[str, Any]]:
+        if subset == "unquantized":
+            return {
+                "float32": {"torch_dtype": "float32", "quant_scheme": None, "quant_config": {}},
+                "float16": {"torch_dtype": "float16", "quant_scheme": None, "quant_config": {}},
+                "bfloat16": {"torch_dtype": "bfloat16", "quant_scheme": None, "quant_config": {}},
+            }
+        else:
+            raise ValueError(f"Unknown subset: {subset}")
+
+    def _get_attention_configs(self) -> List[str]:
+        return ["eager", "sdpa"]
+
+
+if __name__ == "__main__":
+    runner = CPUPyTorchBenchmarkRunner()
+    runner.run_benchmarks()

llm_perf/benchmark_runners/update_llm_perf_cuda_pytorch.py

@@ -0,0 +1,126 @@
+from typing import Any, Dict, List
+
+from llm_perf.common.benchmark_runner import BenchmarkRunner
+from llm_perf.common.utils import GENERATE_KWARGS, INPUT_SHAPES
+from optimum_benchmark import PyTorchConfig
+from optimum_benchmark.benchmark.config import BenchmarkConfig
+from optimum_benchmark.launchers.process.config import ProcessConfig
+from optimum_benchmark.scenarios.inference.config import InferenceConfig
+
+
+class CUDAPyTorchBenchmarkRunner(BenchmarkRunner):
+    def __init__(self):
+        super().__init__(backend="pytorch", hardware="cuda")
+
+    def is_benchmark_supported(self, weights_config: str, attn_implementation: str) -> bool:
+        if attn_implementation == "flash_attention_2" and weights_config == "float32":
+            return False
+        return True
+
+    def get_benchmark_config(self, model: str, attn_implementation: str, weights_config: str) -> BenchmarkConfig:
+        assert (
+            weights_config in self.weights_configs
+        ), f"your config does contains the {weights_config}, adjust your _get_weights_configs to fix this issue"
+
+        torch_dtype = self.weights_configs[weights_config]["torch_dtype"]
+        quant_scheme = self.weights_configs[weights_config]["quant_scheme"]
+        quant_config = self.weights_configs[weights_config]["quant_config"]
+
+        launcher_config = ProcessConfig(device_isolation=True, device_isolation_action="kill")
+        scenario_config = InferenceConfig(
+            memory=True,
+            energy=True,
+            latency=True,
+            duration=10,
+            iterations=10,
+            warmup_runs=10,
+            input_shapes=INPUT_SHAPES,
+            generate_kwargs=GENERATE_KWARGS,
+        )
+        backend_config = PyTorchConfig(
+            model=model,
+            device="cuda",
+            device_ids="0",
+            no_weights=True,
+            library="transformers",
+            task="text-generation",
+            torch_dtype=torch_dtype,
+            quantization_scheme=quant_scheme,
+            quantization_config=quant_config,
+            attn_implementation=attn_implementation,
+            model_kwargs={"trust_remote_code": True},
+        )
+
+        return BenchmarkConfig(
+            name=f"{weights_config}-{attn_implementation}",
+            scenario=scenario_config,
+            launcher=launcher_config,
+            backend=backend_config,
+        )
+
+    def _get_weights_configs(self, subset) -> Dict[str, Dict[str, Any]]:
+        if subset == "unquantized":
+            return {
+                "float32": {"torch_dtype": "float32", "quant_scheme": None, "quant_config": {}},
+                "float16": {"torch_dtype": "float16", "quant_scheme": None, "quant_config": {}},
+                "bfloat16": {"torch_dtype": "bfloat16", "quant_scheme": None, "quant_config": {}},
+            }
+        elif subset == "bnb":
+            return {
+                "4bit-bnb": {"torch_dtype": "float16", "quant_scheme": "bnb", "quant_config": {"load_in_4bit": True}},
+                "8bit-bnb": {"torch_dtype": "float16", "quant_scheme": "bnb", "quant_config": {"load_in_8bit": True}},
+            }
+        elif subset == "gptq":
+            return {
+                "4bit-gptq-exllama-v1": {
+                    "torch_dtype": "float16",
+                    "quant_scheme": "gptq",
+                    "quant_config": {"bits": 4, "use_exllama ": True, "version": 1, "model_seqlen": 256},
+                },
+                "4bit-gptq-exllama-v2": {
+                    "torch_dtype": "float16",
+                    "quant_scheme": "gptq",
+                    "quant_config": {"bits": 4, "use_exllama ": True, "version": 2, "model_seqlen": 256},
+                },
+            }
+        elif subset == "awq":
+            return {
+                "4bit-awq-gemm": {
+                    "torch_dtype": "float16",
+                    "quant_scheme": "awq",
+                    "quant_config": {"bits": 4, "version": "gemm"},
+                },
+                "4bit-awq-gemv": {
+                    "torch_dtype": "float16",
+                    "quant_scheme": "awq",
+                    "quant_config": {"bits": 4, "version": "gemv"},
+                },
+                "4bit-awq-exllama-v1": {
+                    "torch_dtype": "float16",
+                    "quant_scheme": "awq",
+                    "quant_config": {
+                        "bits": 4,
+                        "version": "exllama",
+                        "exllama_config": {"version": 1, "max_input_len": 64, "max_batch_size": 1},
+                    },
+                },
+                "4bit-awq-exllama-v2": {
+                    "torch_dtype": "float16",
+                    "quant_scheme": "awq",
+                    "quant_config": {
+                        "bits": 4,
+                        "version": "exllama",
+                        "exllama_config": {"version": 2, "max_input_len": 64, "max_batch_size": 1},
+                    },
+                },
+            }
+        else:
+            raise ValueError(f"Unknown subset: {subset}")
+
+    def _get_attention_configs(self) -> List[str]:
+        return ["eager", "sdpa", "flash_attention_2"]
+
+
+if __name__ == "__main__":
+    runner = CUDAPyTorchBenchmarkRunner()
+    runner.run_benchmarks()