Fix eagle radix cache

python/sglang/srt/managers/schedule_batch.py

@@ -546,6 +546,8 @@ def __init__(
         self.host_hit_length = 0
         # The node to lock until for swa radix tree lock ref
         self.swa_uuid_for_lock: Optional[int] = None
+        # The prefix length of the last prefix matching
+        self.last_matched_prefix_len: int = 0
 
         # Whether or not if it is chunked. It increments whenever
         # it is chunked, and decrement whenever chunked request is
@@ -700,6 +702,7 @@ def init_next_round_input(
                     token_ids=self.adjust_max_prefix_ids(), extra_key=self.extra_key
                 ),
             )
+            self.last_matched_prefix_len = len(self.prefix_indices)
         self.extend_input_len = len(self.fill_ids) - len(self.prefix_indices)
 
     def adjust_max_prefix_ids(self):

python/sglang/srt/managers/scheduler.py

@@ -756,6 +756,7 @@ def init_memory_pool_and_cache(self):
                     disable=server_args.disable_radix_cache,
                     enable_kv_cache_events=self.enable_kv_cache_events,
                     eviction_policy=server_args.radix_eviction_policy,
+                    is_eagle=self.spec_algorithm.is_eagle(),
                 )
 
         if (

python/sglang/srt/mem_cache/radix_cache.py

@@ -23,7 +23,7 @@
 import time
 from collections import defaultdict
 from functools import partial
-from typing import TYPE_CHECKING, Any, Iterator, List, Optional, Union
+from typing import TYPE_CHECKING, Any, Iterator, List, Optional, Tuple, Union
 
 import torch
 
@@ -159,6 +159,16 @@ def get_child_key(key: RadixKey, page_size: int = 1):
         return (key.extra_key, plain_key)
 
 
+def _convert_to_bigram_key(tokens: List[int]) -> List[Tuple[int, int]]:
+    # EAGLE uses bigram keys in the radix tree since draft sequence is the one-token-shifted version of target
+    # [1, 2, 3, 4] -> [(1,2), (2,3), (3,4)]
+    if len(tokens) < 2:
+        return []
+    if isinstance(tokens[0], tuple):
+        return tokens
+    return [(tokens[i], tokens[i + 1]) for i in range(len(tokens) - 1)]
+
+
 class RadixCache(BasePrefixCache):
     def __init__(
         self,
@@ -168,13 +178,15 @@ def __init__(
         disable: bool = False,
         enable_kv_cache_events: bool = False,
         eviction_policy: str = "lru",
+        is_eagle: bool = False,
     ):
         self.req_to_token_pool = req_to_token_pool
         self.token_to_kv_pool_allocator = token_to_kv_pool_allocator
         self.page_size = page_size
         self.disable = disable
         self.enable_kv_cache_events = enable_kv_cache_events
         self.kv_event_queue = []
+        self.is_eagle = is_eagle
 
         if self.token_to_kv_pool_allocator:
             self.device = self.token_to_kv_pool_allocator.device
@@ -188,6 +200,11 @@ def __init__(
             self.key_match_fn = partial(_key_match_paged, page_size=page_size)
             self.get_child_key_fn = partial(get_child_key, page_size=page_size)
 
+        if is_eagle:
+            self.key_convert_fn = _convert_to_bigram_key
+        else:
+            self.key_convert_fn = lambda key: key
+
         if eviction_policy.lower() == "lru":
             self.eviction_strategy: EvictionStrategy = LRUStrategy()
         elif eviction_policy.lower() == "lfu":
@@ -248,6 +265,8 @@ def match_prefix(self, key: RadixKey, **kwargs) -> MatchResult:
                 to expose a precise boundary; this structural refinement improves
                 subsequent match efficiency and does not duplicate data.
         """
+        key.token_ids = self.key_convert_fn(key.token_ids)
+
         if self.disable or len(key) == 0:
             return MatchResult(
                 device_indices=torch.empty(
@@ -278,8 +297,15 @@ def insert(self, key: RadixKey, value=None, chunked=False):
         if self.disable:
             return 0
 
+        key.token_ids = self.key_convert_fn(key.token_ids)
+
         if value is None:
             value = torch.tensor(key.token_ids, dtype=torch.int64)
+
+        if self.is_eagle:
+            # Make sure the value len equal to the EAGLE bigram key len
+            value = value[: len(key)]
+
         return self._insert_helper(self.root_node, key, value)
 
     def cache_finished_req(self, req: Req):
@@ -293,28 +319,39 @@ def cache_finished_req(self, req: Req):
             return
 
         token_ids = (req.origin_input_ids + req.output_ids)[:-1]
+        all_token_len = len(token_ids)
+        actual_kv_len = all_token_len - 1 if self.is_eagle else all_token_len
         kv_indices = self.req_to_token_pool.req_to_token[
-            req.req_pool_idx, : len(token_ids)
+            req.req_pool_idx, :all_token_len
         ]
 
         if self.page_size != 1:
-            page_aligned_len = len(kv_indices) // self.page_size * self.page_size
+            page_aligned_len = actual_kv_len // self.page_size * self.page_size
             page_aligned_kv_indices = kv_indices[:page_aligned_len].to(
                 dtype=torch.int64, copy=True
             )
             self.token_to_kv_pool_allocator.free(kv_indices[page_aligned_len:])
         else:
-            page_aligned_len = len(kv_indices)
+            page_aligned_len = actual_kv_len
             page_aligned_kv_indices = kv_indices.to(dtype=torch.int64, copy=True)
+            if self.is_eagle:
+                self.token_to_kv_pool_allocator.free(kv_indices[page_aligned_len:])
+
+        page_aligned_token_len = (
+            page_aligned_len + 1 if self.is_eagle else page_aligned_len
+        )
+
+        old_prefix_len = len(req.prefix_indices)
+        if self.is_eagle and old_prefix_len > req.last_matched_prefix_len:
+            # prefix_indices attached partial part (for page_size > 1) and one unmatched token (for EAGLE)
+            old_prefix_len -= 1
 
         # Radix Cache takes one ref in memory pool
         new_prefix_len = self.insert(
-            RadixKey(token_ids[:page_aligned_len], req.extra_key),
+            RadixKey(token_ids[:page_aligned_token_len], req.extra_key),
             page_aligned_kv_indices,
         )
-        self.token_to_kv_pool_allocator.free(
-            kv_indices[len(req.prefix_indices) : new_prefix_len]
-        )
+        self.token_to_kv_pool_allocator.free(kv_indices[old_prefix_len:new_prefix_len])
 
         # Remove req slot release the cache lock
         self.req_to_token_pool.free(req.req_pool_idx)
@@ -326,49 +363,73 @@ def cache_unfinished_req(self, req: Req, chunked=False):
             return
 
         token_ids = req.fill_ids
+        all_token_len = len(token_ids)
+        # The actual kv len for EAGLE is len(token_ids), since EAGLE uses bigram key
+        actual_kv_len = all_token_len - 1 if self.is_eagle else all_token_len
         kv_indices = self.req_to_token_pool.req_to_token[
-            req.req_pool_idx, : len(token_ids)
+            req.req_pool_idx, :all_token_len
         ]
 
         if self.page_size != 1:
-            page_aligned_len = len(kv_indices) // self.page_size * self.page_size
+            page_aligned_len = actual_kv_len // self.page_size * self.page_size
             page_aligned_kv_indices = kv_indices[:page_aligned_len].to(
                 dtype=torch.int64, copy=True
             )
         else:
-            page_aligned_len = len(kv_indices)
+            page_aligned_len = actual_kv_len
             page_aligned_kv_indices = kv_indices.to(dtype=torch.int64, copy=True)
-        page_aligned_token_ids = token_ids[:page_aligned_len]
+
+        # For EAGLE, the page_aligned_len is for the bigram key, the normal key len should +1
+        page_aligned_token_len = (
+            page_aligned_len + 1 if self.is_eagle else page_aligned_len
+        )
+        page_aligned_token_ids = token_ids[:page_aligned_token_len]
+
+        old_prefix_len = len(req.prefix_indices)
+        if self.is_eagle and old_prefix_len > req.last_matched_prefix_len:
+            # prefix_indices attached partial part (for page_size > 1) and one unmatched token (for EAGLE)
+            old_prefix_len -= 1
 
         # Radix Cache takes one ref in memory pool
         new_prefix_len = self.insert(
             RadixKey(page_aligned_token_ids, req.extra_key),
             page_aligned_kv_indices,
             chunked=chunked,
         )
-        self.token_to_kv_pool_allocator.free(
-            kv_indices[len(req.prefix_indices) : new_prefix_len]
-        )
+        self.token_to_kv_pool_allocator.free(kv_indices[old_prefix_len:new_prefix_len])
 
         # The prefix indices could be updated, reuse it
         new_indices, new_last_node, _, _ = self.match_prefix(
             RadixKey(token_ids=page_aligned_token_ids, extra_key=req.extra_key)
         )
         self.req_to_token_pool.write(
-            (req.req_pool_idx, slice(len(req.prefix_indices), len(new_indices))),
-            new_indices[len(req.prefix_indices) :],
+            (req.req_pool_idx, slice(old_prefix_len, len(new_indices))),
+            new_indices[old_prefix_len:],
         )
 
+        # The last_matched_prefix_len is not always equal to len(req.prefix_indices)
+        # since for page_size > 1, the partial part is added to req.prefix_indices, but that part of kv indices is not added to the tree.
+        # It should be freed in the next cache_unfinished_req and final cache_finished_req to avoid memory leak.
+        # So we introduce this `last_matched_prefix_len` field to make sure the partial part can be freed correctly.
+        req.last_matched_prefix_len = len(new_indices)
+
         self.dec_lock_ref(req.last_node)
         self.inc_lock_ref(new_last_node)
 
         # `req.prefix_indices` will be used in `PrefillAdder::add_chunked_req` later
         if self.page_size != 1:
+            # Handle partial page, the partial part should be freed in the next cache_unfinished_req and final cache_finished_req.
             req.prefix_indices = torch.cat(
                 [new_indices, kv_indices[len(new_indices) :]]
             )
         else:
-            req.prefix_indices = new_indices
+            if self.is_eagle:
+                # Attach the kv index of the last token for EAGLE, it can be used in chunked prefill
+                req.prefix_indices = torch.cat(
+                    [new_indices, kv_indices[actual_kv_len:]]
+                )
+            else:
+                req.prefix_indices = new_indices
         req.last_node = new_last_node
 
     def pretty_print(self):

python/sglang/test/test_utils.py

@@ -77,7 +77,8 @@
 # EAGLE
 DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST = "meta-llama/Llama-2-7b-chat-hf"
 DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST = "lmsys/sglang-EAGLE-llama2-chat-7B"
-DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3 = "jamesliu1/sglang-EAGLE3-Llama-3.1-Instruct-8B"
+DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST_EAGLE3 = "meta-llama/Llama-3.1-8B-Instruct"
+DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3 = "lmsys/sglang-EAGLE3-LLaMA3.1-Instruct-8B"
 DEFAULT_STANDALONE_SPECULATIVE_TARGET_MODEL_FOR_TEST = (
     "meta-llama/Llama-3.1-8B-Instruct"
 )

test/srt/test_eagle_infer_a.py

@@ -9,6 +9,8 @@
 from sglang.test.test_utils import (
     DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
     DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
+    DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST_EAGLE3,
+    DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3,
     DEFAULT_MODEL_NAME_FOR_TEST_MLA,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
@@ -35,6 +37,11 @@ class TestEAGLEEngine(CustomTestCase):
     }
     NUM_CONFIGS = 2
 
+    THRESHOLDS = {
+        "batch_avg_accept_len": 1.9,
+        "accept_len": 3.6,
+    }
+
     def setUp(self):
         self.prompt = "Today is a sunny day and I like"
         self.sampling_params = {"temperature": 0, "max_new_tokens": 8}
@@ -63,6 +70,7 @@ def test_correctness(self):
                     self._test_eos_token(engine)
                     self._test_acc_length(engine)
                 finally:
+                    engine.flush_cache()  # check engine alive
                     engine.shutdown()
                 print("=" * 100)
 
@@ -92,7 +100,9 @@ def _test_batch_generation(self, engine):
             "avg_spec_accept_length"
         ]
         print(f"{avg_spec_accept_length=}")
-        self.assertGreater(avg_spec_accept_length, 1.9)
+        self.assertGreater(
+            avg_spec_accept_length, self.THRESHOLDS["batch_avg_accept_len"]
+        )
 
     def _test_eos_token(self, engine):
         prompt = "[INST] <<SYS>>\nYou are a helpful assistant.\n<</SYS>>\nToday is a sunny day and I like [/INST]"
@@ -131,10 +141,7 @@ def _test_acc_length(self, engine):
         )
         print(f"{acc_length=:.4f}, {speed=}")
 
-        if engine.server_args.model_path == DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST:
-            self.assertGreater(acc_length, 3.6)
-        else:
-            self.assertGreater(acc_length, 2.5)
+        self.assertGreater(acc_length, self.THRESHOLDS["accept_len"])
 
 
 class TestEAGLEEngineTokenMap(TestEAGLEEngine):
@@ -151,12 +158,16 @@ class TestEAGLEEngineTokenMap(TestEAGLEEngine):
         "dtype": "float16",
     }
     NUM_CONFIGS = 1
+    THRESHOLDS = {
+        "batch_avg_accept_len": 1.9,
+        "accept_len": 2.5,
+    }
 
 
 class TestEAGLE3Engine(TestEAGLEEngine):
     BASE_CONFIG = {
-        "model_path": "meta-llama/Llama-3.1-8B-Instruct",
-        "speculative_draft_model_path": "jamesliu1/sglang-EAGLE3-Llama-3.1-Instruct-8B",
+        "model_path": DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST_EAGLE3,
+        "speculative_draft_model_path": DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3,
         "speculative_algorithm": "EAGLE3",
         "speculative_num_steps": 5,
         "speculative_eagle_topk": 16,
@@ -166,6 +177,72 @@ class TestEAGLE3Engine(TestEAGLEEngine):
         "dtype": "float16",
     }
     NUM_CONFIGS = 1
+    THRESHOLDS = {
+        "batch_avg_accept_len": 1.75,
+        "accept_len": 3.1,
+    }
+
+
+class TestEAGLERadixCache(CustomTestCase):
+    BASE_CONFIG = {
+        "model_path": DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST_EAGLE3,
+        "speculative_draft_model_path": DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3,
+        "speculative_algorithm": "EAGLE3",
+        "speculative_num_steps": 2,
+        "speculative_eagle_topk": 1,
+        "speculative_num_draft_tokens": 3,
+        "mem_fraction_static": 0.7,
+        "cuda_graph_max_bs": 5,
+        "dtype": "float16",
+    }
+
+    def test_correctness(self):
+        configs = [
+            # Basic config
+            self.BASE_CONFIG,
+            # Chunked prefill
+            {**self.BASE_CONFIG, "chunked_prefill_size": 64},
+            # Chunked prefill & Page Size > 1
+            {**self.BASE_CONFIG, "chunked_prefill_size": 64, "page_size": 4},
+        ]
+
+        for i, config in enumerate(configs):
+            with self.subTest(i=i):
+                print(f"{config=}")
+                engine = sgl.Engine(**config, log_level="info", decode_log_interval=10)
+                try:
+                    self._test_acc_length(engine)
+                finally:
+                    engine.shutdown()
+                print("=" * 100)
+
+    def _test_acc_length(self, engine):
+        warmup_prompt = [
+            "Human: Give me a fully functional FastAPI server. Show the python code.\n\nAssistant:",
+        ]
+        sampling_params = {"temperature": 0, "max_new_tokens": 512}
+        output = engine.generate(warmup_prompt, sampling_params)
+        test_prompt = [
+            "<|start_header_id|>system<|end_header_id|>\n\nYou are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe.  Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.\n\nIf a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nGive me a fully functional FastAPI server. Show the python code.<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+        ]
+        output = engine.generate(test_prompt, sampling_params)
+        output = output[0]
+
+        if "spec_verify_ct" in output["meta_info"]:
+            acc_length = (
+                output["meta_info"]["completion_tokens"]
+                / output["meta_info"]["spec_verify_ct"]
+            )
+        else:
+            acc_length = 1.0
+
+        speed = (
+            output["meta_info"]["completion_tokens"]
+            / output["meta_info"]["e2e_latency"]
+        )
+        print(f"{acc_length=:.4f}, {speed=}")
+
+        self.assertGreater(acc_length, 2.5)
 
 
 @unittest.skipIf(is_in_ci(), "To reduce the CI execution time.")