add quantization_weights_path for fp8 weights

ROCm_performance.md

@@ -37,7 +37,7 @@ For more details, please refer to Quark's documentation.
 
 To use ammo, please follow this [instruction](https://github.com/ROCm/vllm/tree/main/examples/fp8/quantizer), and set `VLLM_FP8_USE_AMMO=1`.
 
-Both quantizers generate a safetensor file that contains the quantized weights and the corresponding scaling factors of your model. The safetensor file should be placed under your model folder. Then we can run a model with fp8 quantization using vllm. When creating `vllm.LLM` object, two additional parameters should be added: `quantization="fp8"` and `quantization_param_path={relative path of the safetensors with your model path}`.
+Both quantizers generate a safetensor file that contains the quantized weights and the corresponding scaling factors of your model. The safetensor file should be placed under your model folder. Then we can run a model with fp8 quantization using vllm. When creating `vllm.LLM` object, two additional parameters should be added: `quantization="fp8"` and `quantization_weights_path={relative path of the safetensors with your model path}`.
 
 ## Gemm Tuning for Fp8
 

benchmarks/benchmark_latency.py

@@ -24,6 +24,7 @@ def main(args: argparse.Namespace):
               num_speculative_tokens=args.num_speculative_tokens,
               tokenizer=args.tokenizer,
               quantization=args.quantization,
+              quantization_weights_path=args.quantization_weights_path,
               tensor_parallel_size=args.tensor_parallel_size,
               trust_remote_code=args.trust_remote_code,
               dtype=args.dtype,
@@ -175,6 +176,13 @@ def run_to_completion(profile_dir: Optional[str] = None):
         'accuracy issues. FP8_E5M2 (without scaling) is only supported on '
         'cuda version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is '
         'instead supported for common inference criteria.')
+    parser.add_argument(
+        '--quantization-weights-path',
+        type=str,
+        default=None,
+        help='Path to the safetensor file containing the quantized weights '
+        'and scaling factors. This should generally be supplied, when '
+        'quantization is FP8.')
     parser.add_argument(
         '--profile',
         action='store_true',

benchmarks/benchmark_throughput.py

@@ -64,6 +64,7 @@ def run_vllm(
     model: str,
     tokenizer: str,
     quantization: Optional[str],
+    quantization_weights_path: Optional[str],
     tensor_parallel_size: int,
     seed: int,
     n: int,
@@ -87,6 +88,7 @@ def run_vllm(
         model=model,
         tokenizer=tokenizer,
         quantization=quantization,
+        quantization_weights_path=quantization_weights_path,
         tensor_parallel_size=tensor_parallel_size,
         seed=seed,
         trust_remote_code=trust_remote_code,
@@ -222,6 +224,7 @@ def main(args: argparse.Namespace):
     if args.backend == "vllm":
         elapsed_time = run_vllm(
             requests, args.model, args.tokenizer, args.quantization,
+            args.quantization_weights_path,
             args.tensor_parallel_size, args.seed, args.n, args.use_beam_search,
             args.trust_remote_code, args.dtype, args.max_model_len,
             args.enforce_eager, args.kv_cache_dtype,
@@ -342,6 +345,13 @@ def main(args: argparse.Namespace):
         'accuracy issues. FP8_E5M2 (without scaling) is only supported on '
         'cuda version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is '
         'instead supported for common inference criteria.')
+    parser.add_argument(
+        '--quantization-weights-path',
+        type=str,
+        default=None,
+        help='Path to the safetensor file containing the quantized weights '
+        'and scaling factors. This should generally be supplied, when '
+        'quantization is FP8.')
     parser.add_argument(
         "--device",
         type=str,

vllm/config.py

@@ -97,6 +97,7 @@ def __init__(
         max_model_len: Optional[int] = None,
         quantization: Optional[str] = None,
         quantization_param_path: Optional[str] = None,
+        quantization_weights_path: Optional[str] = None,
         enforce_eager: bool = False,
         max_context_len_to_capture: Optional[int] = None,
         max_seq_len_to_capture: Optional[int] = None,
@@ -116,6 +117,7 @@ def __init__(
         self.tokenizer_revision = tokenizer_revision
         self.quantization = quantization
         self.quantization_param_path = quantization_param_path
+        self.quantization_weights_path = quantization_weights_path
         self.enforce_eager = enforce_eager
         self.max_context_len_to_capture = max_context_len_to_capture
         if self.max_context_len_to_capture is not None:

vllm/engine/arg_utils.py

@@ -54,6 +54,7 @@ class EngineArgs:
     rope_scaling: Optional[dict] = None
     tokenizer_revision: Optional[str] = None
     quantization: Optional[str] = None
+    quantization_weights_path: Optional[str] = None
     enforce_eager: bool = False
     max_context_len_to_capture: Optional[int] = None
     max_seq_len_to_capture: int = 32768
@@ -337,6 +338,13 @@ def add_cli_args(
                             'None, we assume the model weights are not '
                             'quantized and use `dtype` to determine the data '
                             'type of the weights.')
+        parser.add_argument(
+            '--quantization-weights-path',
+            type=nullable_str,
+            default=None,
+            help='Path to the safetensor file containing the quantized weights '
+            'and scaling factors. This should generally be supplied, when '
+            'quantization is FP8.')
         parser.add_argument('--rope-scaling',
                             default=None,
                             type=json.loads,
@@ -562,7 +570,8 @@ def create_engine_config(self, ) -> EngineConfig:
             self.trust_remote_code, self.dtype, self.seed, self.revision,
             self.code_revision, self.rope_scaling, self.tokenizer_revision,
             self.max_model_len, self.quantization,
-            self.quantization_param_path, self.enforce_eager,
+            self.quantization_param_path, self.quantization_weights_path,
+            self.enforce_eager,
             self.max_context_len_to_capture, self.max_seq_len_to_capture,
             self.max_logprobs, self.disable_sliding_window,
             self.skip_tokenizer_init, self.served_model_name)

vllm/model_executor/model_loader/loader.py

@@ -248,11 +248,11 @@ def load_model(self, *, model_config: ModelConfig,
                                                "fall_back_to_pt_during_load",
                                                True)), )
             if (model_config.quantization == 'fp8'
-                    and model_config.quantization_param_path is not None):
+                    and model_config.quantization_weights_path is not None):
                 model.load_quantized_weights(
                     safetensors_weights_iterator([
                         model_config.model +
-                        model_config.quantization_param_path
+                        model_config.quantization_weights_path
                     ]))
             for _, module in model.named_modules():
                 quant_method = getattr(module, "quant_method", None)