Fix oom issues with fp8 for llama (#1454)

.github/workflows/pr-test.yml

@@ -144,17 +144,17 @@ jobs:
           cd test/srt
           python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_without_radix_cache
 
-      - name: Benchmark Offline Throughput (w/o ChunkedPrefill)
+      - name: Benchmark Offline Throughput (w/ Triton)
         timeout-minutes: 10
         run: |
           cd test/srt
-          python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_without_chunked_prefill
+          python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_with_triton_attention_backend
 
-      - name: Benchmark Offline Throughput (w/ Triton)
+      - name: Benchmark Offline Throughput (w/ FP8)
         timeout-minutes: 10
         run: |
           cd test/srt
-          python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_with_triton_attention_backend
+          python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_default_fp8
 
   performance-test-2-gpu:
     if: github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request'

python/sglang/srt/models/llama.py

@@ -305,8 +305,6 @@ class LlamaForCausalLM(nn.Module):
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
 
-        self.param_dict = dict(self.named_parameters())
-
     @torch.no_grad()
     def forward(
         self,
@@ -374,7 +372,7 @@ class LlamaForCausalLM(nn.Module):
             (".gate_up_proj", ".gate_proj", 0),
             (".gate_up_proj", ".up_proj", 1),
         ]
-        params_dict = self.param_dict
+        params_dict = dict(self.named_parameters())
 
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name or "projector" in name:

python/sglang/srt/models/llama_classification.py

@@ -36,6 +36,7 @@ class LlamaForClassification(nn.Module):
     ) -> None:
         super().__init__()
         self.config = config
+        self.torchao_config = None
         self.quant_config = quant_config
         self.model = LlamaModel(config, quant_config=quant_config)
 
@@ -44,8 +45,6 @@ class LlamaForClassification(nn.Module):
         )
         self.eos_token_id = config.eos_token_id
 
-        self.param_dict = dict(self.named_parameters())
-
     @torch.no_grad()
     def forward(
         self,
@@ -77,7 +76,7 @@ class LlamaForClassification(nn.Module):
         return logits_output
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
-        params_dict = self.param_dict
+        params_dict = dict(self.named_parameters())
 
         for name, loaded_weight in weights:
             if "classification_head" in name:

python/sglang/srt/models/xverse.py

@@ -307,8 +307,6 @@ class XverseForCausalLM(nn.Module):
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
 
-        self.param_dict = dict(self.named_parameters())
-
     @torch.no_grad()
     def forward(
         self,
@@ -333,7 +331,7 @@ class XverseForCausalLM(nn.Module):
             ("gate_up_proj", "gate_proj", 0),
             ("gate_up_proj", "up_proj", 1),
         ]
-        params_dict = self.param_dict
+        params_dict = dict(self.named_parameters())
 
         def load_weights_per_param(name, loaded_weight):
             if "rotary_emb.inv_freq" in name or "projector" in name:

python/sglang/srt/models/xverse_moe.py

@@ -383,8 +383,6 @@ class XverseMoeForCausalLM(nn.Module):
         )
         self.logits_processor = LogitsProcessor(config)
 
-        self.param_dict = dict(self.named_parameters())
-
     @torch.no_grad()
     def forward(
         self,
@@ -406,8 +404,7 @@ class XverseMoeForCausalLM(nn.Module):
             ("gate_up_proj", "gate_proj", 0),
             ("gate_up_proj", "up_proj", 1),
         ]
-
-        params_dict = self.param_dict
+        params_dict = dict(self.named_parameters())
 
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name:

python/sglang/test/test_utils.py

@@ -22,6 +22,7 @@ from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
 from sglang.srt.utils import kill_child_process
 from sglang.utils import get_exception_traceback
 
+DEFAULT_FP8_MODEL_NAME_FOR_TEST = "neuralmagic/Meta-Llama-3.1-8B-FP8"
 DEFAULT_MODEL_NAME_FOR_TEST = "meta-llama/Meta-Llama-3.1-8B-Instruct"
 DEFAULT_MOE_MODEL_NAME_FOR_TEST = "mistralai/Mixtral-8x7B-Instruct-v0.1"
 DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH = 600

test/srt/test_bench_serving.py

 import unittest
 
 from sglang.test.test_utils import (
+    DEFAULT_FP8_MODEL_NAME_FOR_TEST,
     DEFAULT_MODEL_NAME_FOR_TEST,
     DEFAULT_MOE_MODEL_NAME_FOR_TEST,
     is_in_ci,
@@ -59,6 +60,17 @@ class TestBenchServing(unittest.TestCase):
         if is_in_ci():
             assert res["output_throughput"] > 2600
 
+    def test_offline_throughput_default_fp8(self):
+        res = run_bench_serving(
+            model=DEFAULT_FP8_MODEL_NAME_FOR_TEST,
+            num_prompts=500,
+            request_rate=float("inf"),
+            other_server_args=[],
+        )
+
+        if is_in_ci():
+            assert res["output_throughput"] > 3100
+
     def test_online_latency_default(self):
         res = run_bench_serving(
             model=DEFAULT_MODEL_NAME_FOR_TEST,

test/srt/test_chunked_prefill.py

@@ -12,8 +12,10 @@ from sglang.test.test_utils import (
 
 
 class TestChunkedPrefill(unittest.TestCase):
-    def run_mmlu(self, disable_radix_cache, enable_mixed_chunk):
-        other_args = ["--chunked-prefill-size", "32"]
+    def run_mmlu(
+        self, disable_radix_cache, enable_mixed_chunk, chunked_prefill_size=32
+    ):
+        other_args = ["--chunked-prefill-size", str(chunked_prefill_size)]
         if disable_radix_cache:
             other_args += ["--disable-radix-cache"]
 
@@ -55,6 +57,11 @@ class TestChunkedPrefill(unittest.TestCase):
     def test_mixed_chunked_prefill_without_radix_cache(self):
         self.run_mmlu(disable_radix_cache=True, enable_mixed_chunk=True)
 
+    def test_no_chunked_prefill(self):
+        self.run_mmlu(
+            disable_radix_cache=False, enable_mixed_chunk=False, chunked_prefill_size=-1
+        )
+
 
 if __name__ == "__main__":
     unittest.main()