sglang 0.5.2rc1__py3-none-any.whl → 0.5.3rc0__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- sglang/bench_one_batch_server.py +10 -1
- sglang/bench_serving.py +257 -29
- sglang/lang/interpreter.py +1 -1
- sglang/srt/configs/__init__.py +4 -0
- sglang/srt/configs/device_config.py +3 -1
- sglang/srt/configs/dots_vlm.py +139 -0
- sglang/srt/configs/internvl.py +6 -0
- sglang/srt/configs/load_config.py +1 -0
- sglang/srt/configs/model_config.py +50 -6
- sglang/srt/configs/qwen3_next.py +326 -0
- sglang/srt/connector/__init__.py +8 -1
- sglang/srt/connector/remote_instance.py +82 -0
- sglang/srt/constrained/base_grammar_backend.py +48 -12
- sglang/srt/constrained/llguidance_backend.py +0 -1
- sglang/srt/constrained/outlines_backend.py +0 -1
- sglang/srt/constrained/xgrammar_backend.py +28 -9
- sglang/srt/custom_op.py +11 -1
- sglang/srt/debug_utils/dump_comparator.py +81 -44
- sglang/srt/debug_utils/dump_loader.py +97 -0
- sglang/srt/debug_utils/dumper.py +11 -3
- sglang/srt/debug_utils/text_comparator.py +73 -11
- sglang/srt/disaggregation/base/conn.py +1 -1
- sglang/srt/disaggregation/common/conn.py +15 -12
- sglang/srt/disaggregation/decode.py +21 -10
- sglang/srt/disaggregation/decode_schedule_batch_mixin.py +4 -1
- sglang/srt/disaggregation/fake/conn.py +1 -1
- sglang/srt/disaggregation/mini_lb.py +6 -445
- sglang/srt/disaggregation/mooncake/conn.py +18 -10
- sglang/srt/disaggregation/nixl/conn.py +180 -16
- sglang/srt/disaggregation/prefill.py +5 -3
- sglang/srt/disaggregation/utils.py +5 -50
- sglang/srt/distributed/parallel_state.py +67 -43
- sglang/srt/entrypoints/engine.py +38 -17
- sglang/srt/entrypoints/grpc_request_manager.py +580 -0
- sglang/srt/entrypoints/grpc_server.py +680 -0
- sglang/srt/entrypoints/http_server.py +88 -53
- sglang/srt/entrypoints/openai/protocol.py +7 -4
- sglang/srt/entrypoints/openai/serving_base.py +46 -3
- sglang/srt/entrypoints/openai/serving_chat.py +39 -19
- sglang/srt/entrypoints/openai/serving_completions.py +15 -4
- sglang/srt/entrypoints/openai/serving_embedding.py +9 -4
- sglang/srt/entrypoints/openai/serving_rerank.py +3 -1
- sglang/srt/entrypoints/openai/serving_responses.py +7 -4
- sglang/srt/entrypoints/openai/serving_score.py +1 -0
- sglang/srt/eplb/eplb_manager.py +2 -2
- sglang/srt/eplb/expert_distribution.py +26 -13
- sglang/srt/eplb/expert_location.py +8 -3
- sglang/srt/eplb/expert_location_updater.py +1 -1
- sglang/srt/function_call/base_format_detector.py +3 -6
- sglang/srt/function_call/ebnf_composer.py +11 -9
- sglang/srt/function_call/function_call_parser.py +6 -0
- sglang/srt/function_call/glm4_moe_detector.py +1 -1
- sglang/srt/function_call/gpt_oss_detector.py +1 -1
- sglang/srt/function_call/qwen3_coder_detector.py +1 -1
- sglang/srt/grpc/__init__.py +1 -0
- sglang/srt/grpc/sglang_scheduler_pb2.py +106 -0
- sglang/srt/grpc/sglang_scheduler_pb2.pyi +427 -0
- sglang/srt/grpc/sglang_scheduler_pb2_grpc.py +236 -0
- sglang/srt/hf_transformers_utils.py +4 -0
- sglang/srt/layers/activation.py +142 -9
- sglang/srt/layers/attention/aiter_backend.py +93 -68
- sglang/srt/layers/attention/ascend_backend.py +11 -4
- sglang/srt/layers/attention/fla/chunk.py +242 -0
- sglang/srt/layers/attention/fla/chunk_delta_h.py +314 -0
- sglang/srt/layers/attention/fla/chunk_o.py +178 -0
- sglang/srt/layers/attention/fla/chunk_scaled_dot_kkt.py +151 -0
- sglang/srt/layers/attention/fla/cumsum.py +300 -0
- sglang/srt/layers/attention/fla/fused_recurrent.py +640 -0
- sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py +232 -0
- sglang/srt/layers/attention/fla/index.py +37 -0
- sglang/srt/layers/attention/fla/l2norm.py +150 -0
- sglang/srt/layers/attention/fla/layernorm_gated.py +326 -0
- sglang/srt/layers/attention/fla/op.py +66 -0
- sglang/srt/layers/attention/fla/solve_tril.py +465 -0
- sglang/srt/layers/attention/fla/utils.py +331 -0
- sglang/srt/layers/attention/fla/wy_fast.py +158 -0
- sglang/srt/layers/attention/flashinfer_backend.py +6 -4
- sglang/srt/layers/attention/flashinfer_mla_backend.py +16 -12
- sglang/srt/layers/attention/hybrid_attn_backend.py +57 -50
- sglang/srt/layers/attention/hybrid_linear_attn_backend.py +602 -0
- sglang/srt/layers/attention/intel_amx_backend.py +3 -0
- sglang/srt/layers/attention/mamba/causal_conv1d.py +128 -0
- sglang/srt/layers/attention/mamba/causal_conv1d_triton.py +1052 -0
- sglang/srt/layers/attention/mamba/mamba.py +64 -0
- sglang/srt/layers/attention/torch_native_backend.py +12 -6
- sglang/srt/layers/attention/triton_backend.py +18 -1
- sglang/srt/layers/attention/trtllm_mla_backend.py +124 -31
- sglang/srt/layers/attention/wave_ops/decode_attention.py +2 -4
- sglang/srt/layers/attention/wave_ops/extend_attention.py +1 -3
- sglang/srt/layers/communicator.py +45 -7
- sglang/srt/layers/dp_attention.py +30 -1
- sglang/srt/layers/layernorm.py +32 -15
- sglang/srt/layers/linear.py +34 -3
- sglang/srt/layers/logits_processor.py +29 -10
- sglang/srt/layers/moe/__init__.py +2 -1
- sglang/srt/layers/moe/cutlass_w4a8_moe.py +3 -3
- sglang/srt/layers/moe/ep_moe/kernels.py +1 -1
- sglang/srt/layers/moe/ep_moe/layer.py +182 -62
- sglang/srt/layers/moe/flashinfer_cutedsl_moe.py +156 -0
- sglang/srt/layers/moe/fused_moe_native.py +5 -3
- sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_1/E=128,N=768,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
- sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=128,N=352,device_name=NVIDIA_RTX_5880_Ada_Generation,dtype=fp8_w8a8.json +146 -0
- sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=256,N=512,device_name=NVIDIA_H20.json +146 -0
- sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/{E=128,N=768,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json → E=257,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json } +29 -29
- sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=128,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
- sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=128,device_name=NVIDIA_H20-3e.json +146 -0
- sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=128,device_name=NVIDIA_H200.json +146 -0
- sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=256,device_name=NVIDIA_H20-3e.json +146 -0
- sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=256,device_name=NVIDIA_H200.json +146 -0
- sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=64,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
- sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=64,device_name=NVIDIA_H200.json +146 -0
- sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +5 -2
- sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_config.py +1 -1
- sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py +23 -20
- sglang/srt/layers/moe/fused_moe_triton/layer.py +61 -59
- sglang/srt/layers/moe/moe_runner/__init__.py +2 -1
- sglang/srt/layers/moe/moe_runner/base.py +274 -1
- sglang/srt/layers/moe/moe_runner/runner.py +80 -0
- sglang/srt/layers/moe/moe_runner/triton.py +448 -0
- sglang/srt/layers/moe/token_dispatcher/__init__.py +16 -4
- sglang/srt/layers/moe/token_dispatcher/{base_dispatcher.py → base.py} +67 -17
- sglang/srt/layers/moe/token_dispatcher/deepep.py +43 -39
- sglang/srt/layers/moe/token_dispatcher/standard.py +44 -2
- sglang/srt/layers/moe/topk.py +30 -9
- sglang/srt/layers/moe/utils.py +12 -7
- sglang/srt/layers/quantization/awq.py +19 -7
- sglang/srt/layers/quantization/base_config.py +11 -6
- sglang/srt/layers/quantization/blockwise_int8.py +38 -27
- sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py +50 -30
- sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py +13 -1
- sglang/srt/layers/quantization/deep_gemm_wrapper/entrypoint.py +27 -0
- sglang/srt/layers/quantization/fp8.py +76 -47
- sglang/srt/layers/quantization/fp8_utils.py +50 -31
- sglang/srt/layers/quantization/gptq.py +25 -17
- sglang/srt/layers/quantization/modelopt_quant.py +182 -49
- sglang/srt/layers/quantization/moe_wna16.py +21 -18
- sglang/srt/layers/quantization/mxfp4.py +68 -41
- sglang/srt/layers/quantization/quark/quark_moe.py +32 -27
- sglang/srt/layers/quantization/quark/schemes/quark_w4a4_mxfp4.py +49 -30
- sglang/srt/layers/quantization/quark/utils.py +97 -0
- sglang/srt/layers/quantization/rocm_mxfp4_utils.py +13 -0
- sglang/srt/layers/quantization/unquant.py +135 -47
- sglang/srt/layers/quantization/w4afp8.py +30 -17
- sglang/srt/layers/quantization/w8a8_fp8.py +35 -20
- sglang/srt/layers/quantization/w8a8_int8.py +76 -38
- sglang/srt/layers/rocm_linear_utils.py +44 -0
- sglang/srt/layers/rotary_embedding.py +0 -18
- sglang/srt/layers/sampler.py +162 -18
- sglang/srt/lora/backend/base_backend.py +50 -8
- sglang/srt/lora/backend/triton_backend.py +90 -2
- sglang/srt/lora/layers.py +32 -0
- sglang/srt/lora/lora.py +4 -1
- sglang/srt/lora/lora_manager.py +35 -112
- sglang/srt/lora/mem_pool.py +24 -10
- sglang/srt/lora/utils.py +18 -9
- sglang/srt/managers/async_dynamic_batch_tokenizer.py +170 -0
- sglang/srt/managers/cache_controller.py +200 -199
- sglang/srt/managers/data_parallel_controller.py +105 -35
- sglang/srt/managers/detokenizer_manager.py +8 -4
- sglang/srt/managers/disagg_service.py +46 -0
- sglang/srt/managers/io_struct.py +199 -12
- sglang/srt/managers/mm_utils.py +1 -0
- sglang/srt/managers/multi_tokenizer_mixin.py +351 -397
- sglang/srt/managers/schedule_batch.py +77 -56
- sglang/srt/managers/schedule_policy.py +4 -3
- sglang/srt/managers/scheduler.py +191 -139
- sglang/srt/managers/scheduler_metrics_mixin.py +116 -9
- sglang/srt/managers/scheduler_output_processor_mixin.py +55 -11
- sglang/srt/managers/scheduler_profiler_mixin.py +1 -1
- sglang/srt/managers/template_manager.py +3 -3
- sglang/srt/managers/tokenizer_communicator_mixin.py +569 -0
- sglang/srt/managers/tokenizer_manager.py +260 -519
- sglang/srt/managers/tp_worker.py +53 -4
- sglang/srt/managers/tp_worker_overlap_thread.py +42 -19
- sglang/srt/mem_cache/allocator.py +1 -1
- sglang/srt/mem_cache/hicache_storage.py +18 -33
- sglang/srt/mem_cache/hiradix_cache.py +108 -48
- sglang/srt/mem_cache/memory_pool.py +347 -48
- sglang/srt/mem_cache/memory_pool_host.py +121 -57
- sglang/srt/mem_cache/radix_cache.py +0 -2
- sglang/srt/mem_cache/storage/hf3fs/hf3fs_client.py +164 -0
- sglang/srt/mem_cache/storage/hf3fs/{client_hf3fs.py → hf3fs_usrbio_client.py} +5 -1
- sglang/srt/mem_cache/storage/hf3fs/storage_hf3fs.py +95 -5
- sglang/srt/mem_cache/storage/lmcache/lmc_radix_cache.py +280 -0
- sglang/srt/mem_cache/storage/lmcache/unit_test.py +121 -0
- sglang/srt/mem_cache/storage/mooncake_store/mooncake_store.py +81 -20
- sglang/srt/mem_cache/storage/mooncake_store/test_mooncake_store.py +161 -0
- sglang/srt/mem_cache/swa_radix_cache.py +0 -2
- sglang/srt/metrics/collector.py +502 -77
- sglang/srt/metrics/startup_func_log_and_timer.py +150 -0
- sglang/srt/metrics/utils.py +48 -0
- sglang/srt/model_executor/cpu_graph_runner.py +640 -0
- sglang/srt/model_executor/cuda_graph_runner.py +13 -5
- sglang/srt/model_executor/forward_batch_info.py +75 -19
- sglang/srt/model_executor/model_runner.py +357 -30
- sglang/srt/model_loader/__init__.py +9 -3
- sglang/srt/model_loader/loader.py +128 -4
- sglang/srt/model_loader/weight_utils.py +2 -1
- sglang/srt/models/apertus.py +686 -0
- sglang/srt/models/bailing_moe.py +798 -218
- sglang/srt/models/bailing_moe_nextn.py +168 -0
- sglang/srt/models/deepseek_v2.py +346 -48
- sglang/srt/models/dots_vlm.py +174 -0
- sglang/srt/models/dots_vlm_vit.py +337 -0
- sglang/srt/models/ernie4.py +1 -1
- sglang/srt/models/gemma3n_mm.py +1 -1
- sglang/srt/models/glm4_moe.py +11 -2
- sglang/srt/models/glm4v.py +4 -2
- sglang/srt/models/glm4v_moe.py +3 -0
- sglang/srt/models/gpt_oss.py +1 -1
- sglang/srt/models/internvl.py +28 -0
- sglang/srt/models/llama4.py +9 -0
- sglang/srt/models/llama_eagle3.py +13 -0
- sglang/srt/models/longcat_flash.py +2 -2
- sglang/srt/models/minicpmv.py +165 -3
- sglang/srt/models/mllama4.py +25 -0
- sglang/srt/models/opt.py +637 -0
- sglang/srt/models/qwen2.py +7 -0
- sglang/srt/models/qwen2_5_vl.py +27 -3
- sglang/srt/models/qwen2_moe.py +60 -13
- sglang/srt/models/qwen3.py +8 -2
- sglang/srt/models/qwen3_moe.py +40 -9
- sglang/srt/models/qwen3_next.py +1042 -0
- sglang/srt/models/qwen3_next_mtp.py +112 -0
- sglang/srt/models/step3_vl.py +1 -1
- sglang/srt/models/torch_native_llama.py +1 -1
- sglang/srt/multimodal/processors/dots_vlm.py +99 -0
- sglang/srt/multimodal/processors/glm4v.py +9 -9
- sglang/srt/multimodal/processors/internvl.py +141 -129
- sglang/srt/multimodal/processors/qwen_vl.py +15 -5
- sglang/srt/offloader.py +27 -3
- sglang/srt/{reasoning_parser.py → parser/reasoning_parser.py} +1 -1
- sglang/srt/remote_instance_weight_loader_utils.py +69 -0
- sglang/srt/sampling/sampling_batch_info.py +18 -15
- sglang/srt/server_args.py +355 -37
- sglang/srt/speculative/eagle_draft_cuda_graph_runner.py +5 -0
- sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py +10 -1
- sglang/srt/speculative/eagle_utils.py +0 -2
- sglang/srt/speculative/eagle_worker.py +197 -112
- sglang/srt/speculative/spec_info.py +5 -0
- sglang/srt/speculative/standalone_worker.py +109 -0
- sglang/srt/tracing/trace.py +552 -0
- sglang/srt/utils.py +46 -3
- sglang/srt/weight_sync/utils.py +1 -1
- sglang/test/attention/test_trtllm_mla_backend.py +169 -5
- sglang/test/few_shot_gsm8k.py +1 -0
- sglang/test/runners.py +4 -0
- sglang/test/test_cutlass_moe.py +24 -6
- sglang/test/test_disaggregation_utils.py +66 -0
- sglang/test/test_fp4_moe.py +370 -1
- sglang/test/test_utils.py +28 -1
- sglang/utils.py +12 -0
- sglang/version.py +1 -1
- {sglang-0.5.2rc1.dist-info → sglang-0.5.3rc0.dist-info}/METADATA +59 -123
- {sglang-0.5.2rc1.dist-info → sglang-0.5.3rc0.dist-info}/RECORD +263 -200
- sglang/srt/disaggregation/launch_lb.py +0 -118
- sglang/srt/mem_cache/storage/mooncake_store/unit_test.py +0 -40
- /sglang/srt/{model_parallel.py → layers/model_parallel.py} +0 -0
- /sglang/srt/{code_completion_parser.py → parser/code_completion_parser.py} +0 -0
- /sglang/srt/{conversation.py → parser/conversation.py} +0 -0
- /sglang/srt/{harmony_parser.py → parser/harmony_parser.py} +0 -0
- /sglang/srt/{jinja_template_utils.py → parser/jinja_template_utils.py} +0 -0
- {sglang-0.5.2rc1.dist-info → sglang-0.5.3rc0.dist-info}/WHEEL +0 -0
- {sglang-0.5.2rc1.dist-info → sglang-0.5.3rc0.dist-info}/licenses/LICENSE +0 -0
- {sglang-0.5.2rc1.dist-info → sglang-0.5.3rc0.dist-info}/top_level.txt +0 -0
@@ -0,0 +1,161 @@
|
|
1
|
+
import logging
|
2
|
+
import uuid
|
3
|
+
|
4
|
+
import torch
|
5
|
+
from mooncake_store import MooncakeStore
|
6
|
+
|
7
|
+
from sglang.srt.mem_cache.hicache_storage import HiCacheStorageConfig
|
8
|
+
|
9
|
+
logging.basicConfig(
|
10
|
+
level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s"
|
11
|
+
)
|
12
|
+
logger = logging.getLogger(__name__)
|
13
|
+
|
14
|
+
|
15
|
+
def generate_batch_query_keys(kv_num: int, config: HiCacheStorageConfig):
|
16
|
+
keys = []
|
17
|
+
for _ in range(kv_num):
|
18
|
+
key = "test_" + str(uuid.uuid4())
|
19
|
+
keys.append(key)
|
20
|
+
set_keys = []
|
21
|
+
for key in keys:
|
22
|
+
if config.is_mla_model:
|
23
|
+
set_keys.append(key + "_k")
|
24
|
+
else:
|
25
|
+
set_keys.append(key + f"_{config.tp_rank}_k")
|
26
|
+
set_keys.append(key + f"_{config.tp_rank}_v")
|
27
|
+
get_keys = set_keys
|
28
|
+
exist_keys = keys
|
29
|
+
return set_keys, get_keys, exist_keys
|
30
|
+
|
31
|
+
|
32
|
+
def test_single_operation():
|
33
|
+
"""Test the set API with a single key-value pair."""
|
34
|
+
print("=" * 100)
|
35
|
+
print("Testing single operation")
|
36
|
+
|
37
|
+
buffer_size = 1024 * 1024 * 16 # 16MB
|
38
|
+
value_elements = 1024
|
39
|
+
store = MooncakeStore()
|
40
|
+
buffer = torch.randn(buffer_size, dtype=torch.float32)
|
41
|
+
store.register_buffer(buffer)
|
42
|
+
value_size = value_elements * buffer.element_size()
|
43
|
+
|
44
|
+
key = str(uuid.uuid4())
|
45
|
+
set_slice = buffer[:value_elements]
|
46
|
+
get_slice = buffer[value_elements : 2 * value_elements]
|
47
|
+
set_location = set_slice.data_ptr()
|
48
|
+
get_location = get_slice.data_ptr()
|
49
|
+
|
50
|
+
# Test set operation
|
51
|
+
result = store.set(key, target_location=set_location, target_sizes=value_size)
|
52
|
+
assert result is True, f"❌set operation failed for key: {key}"
|
53
|
+
|
54
|
+
# Test exists operation
|
55
|
+
assert store.exists(key), f"❌key {key} should exist after set operation"
|
56
|
+
|
57
|
+
# Test get operation
|
58
|
+
result = store.get(key, target_location=get_location, target_sizes=value_size)
|
59
|
+
assert result is True, f"❌get operation failed for key: {key}"
|
60
|
+
|
61
|
+
# Compare the data using proper tensor indices
|
62
|
+
assert torch.allclose(
|
63
|
+
set_slice, get_slice, atol=1e-6
|
64
|
+
), f"❌get operation failed for key: {key}"
|
65
|
+
|
66
|
+
logger.info(f"✅ Single operation passed")
|
67
|
+
|
68
|
+
|
69
|
+
def test_batch_operation(config: HiCacheStorageConfig):
|
70
|
+
"""Test the batch set/get APIs with multiple key-value pairs."""
|
71
|
+
print("=" * 100)
|
72
|
+
print(f"Testing batch operation with config: {config}")
|
73
|
+
|
74
|
+
buffer_size = 1024 * 1024 * 16 # 16MB
|
75
|
+
value_elements = 256
|
76
|
+
kv_num = 13
|
77
|
+
store = MooncakeStore(config)
|
78
|
+
buffer = torch.randn(buffer_size, dtype=torch.float32)
|
79
|
+
store.register_buffer(buffer)
|
80
|
+
value_size = value_elements * buffer.element_size()
|
81
|
+
|
82
|
+
set_keys, get_keys, exist_keys = generate_batch_query_keys(kv_num, config)
|
83
|
+
set_slices = [
|
84
|
+
buffer[i * value_elements : (i + 1) * value_elements]
|
85
|
+
for i in range(len(set_keys))
|
86
|
+
]
|
87
|
+
set_locations = [set_slice.data_ptr() for set_slice in set_slices]
|
88
|
+
target_sizes = [value_size for _ in range(len(set_keys))]
|
89
|
+
|
90
|
+
# Test batch set operation
|
91
|
+
result = store.batch_set(
|
92
|
+
set_keys, target_locations=set_locations, target_sizes=target_sizes
|
93
|
+
)
|
94
|
+
assert result is True, f"❌batch set operation failed"
|
95
|
+
|
96
|
+
# Test batch exists operation
|
97
|
+
assert store.batch_exists(
|
98
|
+
exist_keys
|
99
|
+
), f"❌keys should exist after batch set operation"
|
100
|
+
|
101
|
+
# Test batch get operation
|
102
|
+
get_slices = [
|
103
|
+
buffer[
|
104
|
+
(len(set_keys) + i)
|
105
|
+
* value_elements : (len(set_keys) + i + 1)
|
106
|
+
* value_elements
|
107
|
+
]
|
108
|
+
for i in range(len(get_keys))
|
109
|
+
]
|
110
|
+
get_locations = [get_slice.data_ptr() for get_slice in get_slices]
|
111
|
+
result = store.batch_get(
|
112
|
+
get_keys, target_locations=get_locations, target_sizes=target_sizes
|
113
|
+
)
|
114
|
+
assert result == kv_num, f"❌batch get operation failed"
|
115
|
+
for i in range(len(get_keys)):
|
116
|
+
assert torch.allclose(
|
117
|
+
set_slices[i], get_slices[i], atol=1e-6
|
118
|
+
), f"❌batch get operation failed for key: {get_keys[i]}"
|
119
|
+
|
120
|
+
logger.info(f"✅ Batch operation passed")
|
121
|
+
|
122
|
+
|
123
|
+
if __name__ == "__main__":
|
124
|
+
test_single_operation()
|
125
|
+
test_batch_operation(
|
126
|
+
HiCacheStorageConfig(
|
127
|
+
is_mla_model=False,
|
128
|
+
tp_rank=0,
|
129
|
+
tp_size=1,
|
130
|
+
model_name=None,
|
131
|
+
is_page_first_layout=True,
|
132
|
+
)
|
133
|
+
)
|
134
|
+
test_batch_operation(
|
135
|
+
HiCacheStorageConfig(
|
136
|
+
is_mla_model=True,
|
137
|
+
tp_rank=0,
|
138
|
+
tp_size=1,
|
139
|
+
model_name=None,
|
140
|
+
is_page_first_layout=True,
|
141
|
+
)
|
142
|
+
)
|
143
|
+
test_batch_operation(
|
144
|
+
HiCacheStorageConfig(
|
145
|
+
is_mla_model=False,
|
146
|
+
tp_rank=1,
|
147
|
+
tp_size=4,
|
148
|
+
model_name=None,
|
149
|
+
is_page_first_layout=True,
|
150
|
+
)
|
151
|
+
)
|
152
|
+
test_batch_operation(
|
153
|
+
HiCacheStorageConfig(
|
154
|
+
is_mla_model=True,
|
155
|
+
tp_rank=3,
|
156
|
+
tp_size=8,
|
157
|
+
model_name=None,
|
158
|
+
is_page_first_layout=True,
|
159
|
+
)
|
160
|
+
)
|
161
|
+
logger.info(f"✅ All tests passed")
|