| 28 | |
| 29 | |
| 30 | class ScriptConfig(pydra.Config): |
| 31 | def __init__(self): |
| 32 | self.server_config = ServerConfig() |
| 33 | self.sc = pydra.Alias("server_config") |
| 34 | self.server_config.kv_cache_num_tokens = 1024 * 512 |
| 35 | |
| 36 | self.num_dec = 1 |
| 37 | self.dec_len = 1024 |
| 38 | |
| 39 | self.num_pre = 0 |
| 40 | self.pre_len = 1024 |
| 41 | |
| 42 | self.num_hyd = 0 |
| 43 | self.hyd_shared_len = 1024 |
| 44 | self.hyd_unique_len = 32 |
| 45 | |
| 46 | self.dtype = "bfloat16" |
| 47 | self.pp_rank = 0 |
| 48 | self.num_iters = 10 |
| 49 | self.num_warmup = 5 |
| 50 | self.compile = False |
| 51 | self.dynamic = True |
| 52 | self.fullgraph = True |
| 53 | |
| 54 | self.profile = False |
| 55 | self.profile_name = "bench_model" |
| 56 | |
| 57 | self.non_blocking = True |
| 58 | self.plan_before = False |
| 59 | self.only_plan = False |
| 60 | |
| 61 | self.graph = False |
| 62 | |
| 63 | def prof(self): |
| 64 | self.profile = True |
| 65 | self.num_warmup = 3 |
| 66 | self.num_iters = 10 |
| 67 | self.num_profile_repeat = 3 |
| 68 | |
| 69 | def total_tokens(self): |
| 70 | return self.num_dec + self.num_pre * self.pre_len + self.num_hyd |
| 71 | |
| 72 | def finalize(self): |
| 73 | self.server_config.max_tokens_per_forward = self.total_tokens() |
| 74 | self.server_config.max_seqs_per_forward = self.total_tokens() |
| 75 | |
| 76 | if self.graph: |
| 77 | assert self.plan_before and not self.only_plan |
| 78 | |
| 79 | def l8(self): |
| 80 | self.server_config.model = "meta-llama/Llama-3.1-8B-Instruct" |
| 81 | self.server_config.kv_cache_num_tokens = 1024 * 192 |
| 82 | |
| 83 | def l70(self): |
| 84 | self.server_config.model = "meta-llama/Llama-3.1-70B-Instruct" |
| 85 | |
| 86 | def l1(self): |
| 87 | self.server_config.model = "meta-llama/Llama-3.2-1B-Instruct" |
nothing calls this directly
no outgoing calls
no test coverage detected