Add files using upload-large-folder tool

evals/general_suite/gsm8k_teacher_1280/student/results_2026-07-30T18-42-41.884255.json

@@ -0,0 +1,165 @@
+{
+  "results": {
+    "gsm8k_cot_zeroshot": {
+      "name": "gsm8k_cot_zeroshot",
+      "alias": "gsm8k_cot_zeroshot",
+      "sample_len": 1319,
+      "exact_match,strict-match": 0.0,
+      "exact_match_stderr,strict-match": 0.0,
+      "exact_match,flexible-extract": 0.6937073540561031,
+      "exact_match_stderr,flexible-extract": 0.01269693010656291
+    }
+  },
+  "group_subtasks": {},
+  "configs": {
+    "gsm8k_cot_zeroshot": {
+      "task": "gsm8k_cot_zeroshot",
+      "dataset_path": "openai/gsm8k",
+      "dataset_name": "main",
+      "training_split": "train",
+      "test_split": "test",
+      "fewshot_split": "train",
+      "doc_to_text": "Q: {{question}}\nA: Let's think step by step.",
+      "doc_to_target": "{{answer}}",
+      "unsafe_code": false,
+      "description": "",
+      "target_delimiter": " ",
+      "fewshot_delimiter": "\n\n",
+      "fewshot_config": {
+        "sampler": "default",
+        "split": "train",
+        "process_docs": null,
+        "fewshot_indices": null,
+        "samples": null,
+        "doc_to_text": "Q: {{question}}\nA: Let's think step by step.",
+        "doc_to_choice": null,
+        "doc_to_target": "{{answer}}",
+        "gen_prefix": null,
+        "fewshot_delimiter": "\n\n",
+        "target_delimiter": " "
+      },
+      "num_fewshot": 0,
+      "metric_list": [
+        {
+          "metric": "exact_match",
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "ignore_case": true,
+          "ignore_punctuation": false,
+          "regexes_to_ignore": [
+            ",",
+            "\\$",
+            "(?s).*#### ",
+            "\\.$"
+          ]
+        }
+      ],
+      "output_type": "generate_until",
+      "generation_kwargs": {
+        "until": [
+          "Q:",
+          "</s>",
+          "<|im_end|>"
+        ],
+        "do_sample": false,
+        "max_gen_toks": 1280
+      },
+      "repeats": 1,
+      "filter_list": [
+        {
+          "name": "strict-match",
+          "filter": [
+            {
+              "function": "regex",
+              "regex_pattern": "The answer is (\\-?[0-9\\.\\,]+)."
+            },
+            {
+              "function": "take_first"
+            }
+          ]
+        },
+        {
+          "name": "flexible-extract",
+          "filter": [
+            {
+              "function": "regex",
+              "group_select": -1,
+              "regex_pattern": "(-?[$0-9.,]{2,})|(-?[0-9]+)"
+            },
+            {
+              "function": "take_first"
+            }
+          ]
+        }
+      ],
+      "should_decontaminate": false,
+      "metadata": {
+        "version": 3.0,
+        "model": "student",
+        "base_url": "http://127.0.0.1:8399/v1/chat/completions",
+        "num_concurrent": 32,
+        "tokenized_requests": false,
+        "max_retries": 3,
+        "config_source": "/home/henry/Documents/PythonProjects/variable-reap/.venv/lib/python3.12/site-packages/lm_eval/tasks/gsm8k/gsm8k-cot-zeroshot.yaml"
+      }
+    }
+  },
+  "versions": {
+    "gsm8k_cot_zeroshot": 3.0
+  },
+  "n-shot": {
+    "gsm8k_cot_zeroshot": 0
+  },
+  "higher_is_better": {
+    "gsm8k_cot_zeroshot": {
+      "exact_match": true
+    }
+  },
+  "n-samples": {
+    "gsm8k_cot_zeroshot": {
+      "original": 1319,
+      "effective": 1319
+    }
+  },
+  "config": {
+    "model": "local-chat-completions",
+    "model_args": {
+      "model": "student",
+      "base_url": "http://127.0.0.1:8399/v1/chat/completions",
+      "num_concurrent": 32,
+      "tokenized_requests": false,
+      "max_retries": 3
+    },
+    "batch_size": 1,
+    "batch_sizes": [],
+    "device": "cuda:0",
+    "use_cache": null,
+    "limit": null,
+    "bootstrap_iters": 100000,
+    "gen_kwargs": {
+      "max_gen_toks": 1280
+    },
+    "random_seed": 0,
+    "numpy_seed": 1234,
+    "torch_seed": 1234,
+    "fewshot_seed": 1234
+  },
+  "git_hash": "22365a3",
+  "date": 1785462077.3721132,
+  "pretty_env_info": "PyTorch version: 2.12.1+cu126\nIs debug build: False\nCUDA used to build PyTorch: 12.6\nROCM used to build PyTorch: N/A\n\nOS: Pop!_OS 22.04 LTS (x86_64)\nGCC version: (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0\nClang version: Could not collect\nCMake version: version 3.22.1\nLibc version: glibc-2.35\n\nPython version: 3.12.12 (main, Oct 31 2025, 23:02:31) [Clang 21.1.4 ] (64-bit runtime)\nPython platform: Linux-7.0.11-76070011-generic-x86_64-with-glibc2.35\nIs CUDA available: True\nCUDA runtime version: 12.8.93\nCUDA_MODULE_LOADING set to: \nGPU models and configuration: \nGPU 0: NVIDIA GeForce RTX 3090\nGPU 1: NVIDIA GeForce RTX 3090\nGPU 2: NVIDIA GeForce RTX 3090\n\nNvidia driver version: 580.159.03\ncuDNN version: Could not collect\nIs XPU available: False\nHIP runtime version: N/A\nMIOpen runtime version: N/A\nIs XNNPACK available: True\nCaching allocator config: N/A\n\nCPU:\nArchitecture:                            x86_64\nCPU op-mode(s):                          32-bit, 64-bit\nAddress sizes:                           43 bits physical, 48 bits virtual\nByte Order:                              Little Endian\nCPU(s):                                  48\nOn-line CPU(s) list:                     0-47\nVendor ID:                               AuthenticAMD\nModel name:                              AMD EPYC 7402P 24-Core Processor\nCPU family:                              23\nModel:                                   49\nThread(s) per core:                      2\nCore(s) per socket:                      24\nSocket(s):                               1\nStepping:                                0\nFrequency boost:                         enabled\nCPU max MHz:                             3350.8049\nCPU min MHz:                             1500.0000\nBogoMIPS:                                5600.11\nFlags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sev sev_es\nVirtualization:                          AMD-V\nL1d cache:                               768 KiB (24 instances)\nL1i cache:                               768 KiB (24 instances)\nL2 cache:                                12 MiB (24 instances)\nL3 cache:                                128 MiB (8 instances)\nNUMA node(s):                            1\nNUMA node0 CPU(s):                       0-47\nVulnerability Gather data sampling:      Not affected\nVulnerability Ghostwrite:                Not affected\nVulnerability Indirect target selection: Not affected\nVulnerability Itlb multihit:             Not affected\nVulnerability L1tf:                      Not affected\nVulnerability Mds:                       Not affected\nVulnerability Meltdown:                  Not affected\nVulnerability Mmio stale data:           Not affected\nVulnerability Old microcode:             Not affected\nVulnerability Reg file data sampling:    Not affected\nVulnerability Retbleed:                  Mitigation; untrained return thunk; SMT enabled with STIBP protection\nVulnerability Spec rstack overflow:      Mitigation; Safe RET\nVulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl\nVulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization\nVulnerability Spectre v2:                Mitigation; Retpolines; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected\nVulnerability Srbds:                     Not affected\nVulnerability Tsa:                       Not affected\nVulnerability Tsx async abort:           Not affected\nVulnerability Vmscape:                   Mitigation; IBPB before exit to userspace\n\nVersions of relevant libraries:\n[pip3] Could not collect\n[conda] Could not collect",
+  "transformers_version": "4.57.6",
+  "lm_eval_version": "0.4.12",
+  "upper_git_hash": null,
+  "task_hashes": {
+    "gsm8k_cot_zeroshot": "77ad80425a4cb85562b27f164a9c92977f775bb66a0ec4d374e4ee7b4656a7f5"
+  },
+  "model_source": "local-chat-completions",
+  "model_name": "student",
+  "model_name_sanitized": "student",
+  "system_instruction": null,
+  "system_instruction_sha": null,
+  "fewshot_as_multiturn": true,
+  "chat_template": "",
+  "chat_template_sha": null,
+  "total_evaluation_time_seconds": "91.85732138715684"
+}

evals/general_suite/gsm8k_teacher_1280/student/results_2026-07-30T18-45-59.944024.json

@@ -0,0 +1,225 @@
+{
+  "results": {
+    "humaneval": {
+      "name": "humaneval",
+      "alias": "humaneval",
+      "sample_len": 164,
+      "pass@1,create_test": 0.32926829268292684,
+      "pass@1_stderr,create_test": 0.03680913164548253
+    },
+    "mbpp": {
+      "name": "mbpp",
+      "alias": "mbpp",
+      "sample_len": 500,
+      "pass_at_1,none": 0.3,
+      "pass_at_1_stderr,none": 0.020514426225628046
+    }
+  },
+  "group_subtasks": {},
+  "configs": {
+    "humaneval": {
+      "task": "humaneval",
+      "dataset_path": "openai/openai_humaneval",
+      "test_split": "test",
+      "doc_to_text": "{{prompt}}",
+      "doc_to_target": "{{test}}\ncheck({{entry_point}})",
+      "unsafe_code": true,
+      "description": "",
+      "target_delimiter": " ",
+      "fewshot_delimiter": "\n\n",
+      "fewshot_config": {
+        "sampler": "default",
+        "split": null,
+        "process_docs": null,
+        "fewshot_indices": null,
+        "samples": null,
+        "doc_to_text": "{{prompt}}",
+        "doc_to_choice": null,
+        "doc_to_target": "{{test}}\ncheck({{entry_point}})",
+        "gen_prefix": null,
+        "fewshot_delimiter": "\n\n",
+        "target_delimiter": " "
+      },
+      "num_fewshot": 0,
+      "metric_list": [
+        {
+          "metric": "def pass_at_k(references: list[str], predictions: list[list[str]], k: list[int] = None):\n    global compute_\n    assert k is not None\n    if isinstance(k, int):\n        k = [k]\n    res = compute_.compute(\n        references=references,\n        predictions=predictions,\n        k=k,\n    )\n    return res[0]\n",
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "k": [
+            1
+          ]
+        }
+      ],
+      "output_type": "generate_until",
+      "generation_kwargs": {
+        "until": [
+          "\nclass",
+          "\ndef",
+          "\n#",
+          "\nif",
+          "\nprint"
+        ],
+        "max_gen_toks": 1024,
+        "do_sample": false
+      },
+      "repeats": 1,
+      "filter_list": [
+        {
+          "name": "create_test",
+          "filter": [
+            {
+              "function": "custom",
+              "filter_fn": "<function build_predictions at 0x77ef7066e0c0>"
+            }
+          ]
+        }
+      ],
+      "should_decontaminate": false,
+      "metadata": {
+        "version": 1.0,
+        "model": "student",
+        "base_url": "http://127.0.0.1:8399/v1/completions",
+        "tokenizer": "allenai/OLMoE-1B-7B-0125-Instruct",
+        "num_concurrent": 32,
+        "tokenized_requests": false,
+        "max_retries": 3,
+        "config_source": "/home/henry/Documents/PythonProjects/variable-reap/.venv/lib/python3.12/site-packages/lm_eval/tasks/humaneval/humaneval.yaml"
+      }
+    },
+    "mbpp": {
+      "task": "mbpp",
+      "dataset_path": "google-research-datasets/mbpp",
+      "dataset_name": "full",
+      "test_split": "test",
+      "doc_to_text": "You are an expert Python programmer, and here is your task: {{text}} Your code should pass these tests:\n\n{{test_list[0]}}\n{{test_list[1]}}\n{{test_list[2]}}\n[BEGIN]\n",
+      "doc_to_target": "{% if is_fewshot is defined %}{{code}}\n[DONE]{% else %}{{test_list[0]}}\n{{test_list[1]}}\n{{test_list[2]}}{% endif %}",
+      "unsafe_code": true,
+      "description": "",
+      "target_delimiter": "",
+      "fewshot_delimiter": "\n\n",
+      "fewshot_config": {
+        "sampler": "first_n",
+        "split": null,
+        "process_docs": null,
+        "fewshot_indices": null,
+        "samples": "<function list_fewshot_samples at 0x77f039503920>",
+        "doc_to_text": "You are an expert Python programmer, and here is your task: {{text}} Your code should pass these tests:\n\n{{test_list[0]}}\n{{test_list[1]}}\n{{test_list[2]}}\n[BEGIN]\n",
+        "doc_to_choice": null,
+        "doc_to_target": "{% if is_fewshot is defined %}{{code}}\n[DONE]{% else %}{{test_list[0]}}\n{{test_list[1]}}\n{{test_list[2]}}{% endif %}",
+        "gen_prefix": null,
+        "fewshot_delimiter": "\n\n",
+        "target_delimiter": ""
+      },
+      "num_fewshot": 3,
+      "metric_list": [
+        {
+          "metric": "def pass_at_1(\n    references: Union[str, list[str]], predictions: Union[str, list[list[str]]]\n) -> float:\n    if isinstance(references, str):\n        references = [references]\n    if isinstance(predictions[0], str):\n        predictions = [[p] for p in predictions]\n    return pass_at_k.compute(\n        references=references,\n        predictions=predictions,\n        k=[1],\n    )[0][\"pass@1\"]\n",
+          "aggregation": "mean",
+          "higher_is_better": true
+        }
+      ],
+      "output_type": "generate_until",
+      "generation_kwargs": {
+        "until": [
+          "[DONE]"
+        ],
+        "do_sample": false
+      },
+      "repeats": 1,
+      "should_decontaminate": false,
+      "metadata": {
+        "version": 1.0,
+        "model": "student",
+        "base_url": "http://127.0.0.1:8399/v1/completions",
+        "tokenizer": "allenai/OLMoE-1B-7B-0125-Instruct",
+        "num_concurrent": 32,
+        "tokenized_requests": false,
+        "max_retries": 3,
+        "config_source": "/home/henry/Documents/PythonProjects/variable-reap/.venv/lib/python3.12/site-packages/lm_eval/tasks/mbpp/mbpp.yaml"
+      }
+    }
+  },
+  "versions": {
+    "humaneval": 1.0,
+    "mbpp": 1.0
+  },
+  "n-shot": {
+    "humaneval": 0,
+    "mbpp": 3
+  },
+  "higher_is_better": {
+    "humaneval": {
+      "pass_at_k": true,
+      "pass@1": true
+    },
+    "mbpp": {
+      "pass_at_1": true
+    }
+  },
+  "n-samples": {
+    "humaneval": {
+      "original": 164,
+      "effective": 164
+    },
+    "mbpp": {
+      "original": 500,
+      "effective": 500
+    }
+  },
+  "config": {
+    "model": "local-completions",
+    "model_args": {
+      "model": "student",
+      "base_url": "http://127.0.0.1:8399/v1/completions",
+      "tokenizer": "allenai/OLMoE-1B-7B-0125-Instruct",
+      "num_concurrent": 32,
+      "tokenized_requests": false,
+      "max_retries": 3
+    },
+    "batch_size": 1,
+    "batch_sizes": [],
+    "device": "cuda:0",
+    "use_cache": null,
+    "limit": null,
+    "bootstrap_iters": 100000,
+    "gen_kwargs": {},
+    "random_seed": 0,
+    "numpy_seed": 1234,
+    "torch_seed": 1234,
+    "fewshot_seed": 1234
+  },
+  "git_hash": "22365a3",
+  "date": 1785462170.3520582,
+  "pretty_env_info": "PyTorch version: 2.12.1+cu126\nIs debug build: False\nCUDA used to build PyTorch: 12.6\nROCM used to build PyTorch: N/A\n\nOS: Pop!_OS 22.04 LTS (x86_64)\nGCC version: (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0\nClang version: Could not collect\nCMake version: version 3.22.1\nLibc version: glibc-2.35\n\nPython version: 3.12.12 (main, Oct 31 2025, 23:02:31) [Clang 21.1.4 ] (64-bit runtime)\nPython platform: Linux-7.0.11-76070011-generic-x86_64-with-glibc2.35\nIs CUDA available: True\nCUDA runtime version: 12.8.93\nCUDA_MODULE_LOADING set to: \nGPU models and configuration: \nGPU 0: NVIDIA GeForce RTX 3090\nGPU 1: NVIDIA GeForce RTX 3090\nGPU 2: NVIDIA GeForce RTX 3090\n\nNvidia driver version: 580.159.03\ncuDNN version: Could not collect\nIs XPU available: False\nHIP runtime version: N/A\nMIOpen runtime version: N/A\nIs XNNPACK available: True\nCaching allocator config: N/A\n\nCPU:\nArchitecture:                            x86_64\nCPU op-mode(s):                          32-bit, 64-bit\nAddress sizes:                           43 bits physical, 48 bits virtual\nByte Order:                              Little Endian\nCPU(s):                                  48\nOn-line CPU(s) list:                     0-47\nVendor ID:                               AuthenticAMD\nModel name:                              AMD EPYC 7402P 24-Core Processor\nCPU family:                              23\nModel:                                   49\nThread(s) per core:                      2\nCore(s) per socket:                      24\nSocket(s):                               1\nStepping:                                0\nFrequency boost:                         enabled\nCPU max MHz:                             3350.8049\nCPU min MHz:                             1500.0000\nBogoMIPS:                                5600.11\nFlags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sev sev_es\nVirtualization:                          AMD-V\nL1d cache:                               768 KiB (24 instances)\nL1i cache:                               768 KiB (24 instances)\nL2 cache:                                12 MiB (24 instances)\nL3 cache:                                128 MiB (8 instances)\nNUMA node(s):                            1\nNUMA node0 CPU(s):                       0-47\nVulnerability Gather data sampling:      Not affected\nVulnerability Ghostwrite:                Not affected\nVulnerability Indirect target selection: Not affected\nVulnerability Itlb multihit:             Not affected\nVulnerability L1tf:                      Not affected\nVulnerability Mds:                       Not affected\nVulnerability Meltdown:                  Not affected\nVulnerability Mmio stale data:           Not affected\nVulnerability Old microcode:             Not affected\nVulnerability Reg file data sampling:    Not affected\nVulnerability Retbleed:                  Mitigation; untrained return thunk; SMT enabled with STIBP protection\nVulnerability Spec rstack overflow:      Mitigation; Safe RET\nVulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl\nVulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization\nVulnerability Spectre v2:                Mitigation; Retpolines; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected\nVulnerability Srbds:                     Not affected\nVulnerability Tsa:                       Not affected\nVulnerability Tsx async abort:           Not affected\nVulnerability Vmscape:                   Mitigation; IBPB before exit to userspace\n\nVersions of relevant libraries:\n[pip3] Could not collect\n[conda] Could not collect",
+  "transformers_version": "4.57.6",
+  "lm_eval_version": "0.4.12",
+  "upper_git_hash": null,
+  "tokenizer_pad_token": [
+    "<pad>",
+    "50280"
+  ],
+  "tokenizer_eos_token": [
+    "|||IP_ADDRESS|||",
+    "50279"
+  ],
+  "tokenizer_bos_token": [
+    "|||IP_ADDRESS|||",
+    "50279"
+  ],
+  "eot_token_id": 50279,
+  "max_length": 2047,
+  "task_hashes": {
+    "humaneval": "c122632f0bdffdd0162cc8033a879dfe9d3b35e82c52eda6a767069e0a2beb40",
+    "mbpp": "62356f643ed511c895814c093ddb55451941156ba9febb31dc87c6e00d9723e3"
+  },
+  "model_source": "local-completions",
+  "model_name": "student",
+  "model_name_sanitized": "student",
+  "system_instruction": null,
+  "system_instruction_sha": null,
+  "fewshot_as_multiturn": null,
+  "chat_template": null,
+  "chat_template_sha": null,
+  "total_evaluation_time_seconds": "195.98191328300163"
+}

evals/general_suite/policy_confirm/combo_on25_seed1226/student/results_2026-07-20T09-31-30.699191.json

@@ -0,0 +1,225 @@
+{
+  "results": {
+    "humaneval": {
+      "name": "humaneval",
+      "alias": "humaneval",
+      "sample_len": 164,
+      "pass@1,create_test": 0.15853658536585366,
+      "pass@1_stderr,create_test": 0.028608070471683646
+    },
+    "mbpp": {
+      "name": "mbpp",
+      "alias": "mbpp",
+      "sample_len": 500,
+      "pass_at_1,none": 0.148,
+      "pass_at_1_stderr,none": 0.015896458561251242
+    }
+  },
+  "group_subtasks": {},
+  "configs": {
+    "humaneval": {
+      "task": "humaneval",
+      "dataset_path": "openai/openai_humaneval",
+      "test_split": "test",
+      "doc_to_text": "{{prompt}}",
+      "doc_to_target": "{{test}}\ncheck({{entry_point}})",
+      "unsafe_code": true,
+      "description": "",
+      "target_delimiter": " ",
+      "fewshot_delimiter": "\n\n",
+      "fewshot_config": {
+        "sampler": "default",
+        "split": null,
+        "process_docs": null,
+        "fewshot_indices": null,
+        "samples": null,
+        "doc_to_text": "{{prompt}}",
+        "doc_to_choice": null,
+        "doc_to_target": "{{test}}\ncheck({{entry_point}})",
+        "gen_prefix": null,
+        "fewshot_delimiter": "\n\n",
+        "target_delimiter": " "
+      },
+      "num_fewshot": 0,
+      "metric_list": [
+        {
+          "metric": "def pass_at_k(references: list[str], predictions: list[list[str]], k: list[int] = None):\n    global compute_\n    assert k is not None\n    if isinstance(k, int):\n        k = [k]\n    res = compute_.compute(\n        references=references,\n        predictions=predictions,\n        k=k,\n    )\n    return res[0]\n",
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "k": [
+            1
+          ]
+        }
+      ],
+      "output_type": "generate_until",
+      "generation_kwargs": {
+        "until": [
+          "\nclass",
+          "\ndef",
+          "\n#",
+          "\nif",
+          "\nprint"
+        ],
+        "max_gen_toks": 1024,
+        "do_sample": false
+      },
+      "repeats": 1,
+      "filter_list": [
+        {
+          "name": "create_test",
+          "filter": [
+            {
+              "function": "custom",
+              "filter_fn": "<function build_predictions at 0x75db7fb6af20>"
+            }
+          ]
+        }
+      ],
+      "should_decontaminate": false,
+      "metadata": {
+        "version": 1.0,
+        "model": "student",
+        "base_url": "http://127.0.0.1:8432/v1/completions",
+        "tokenizer": "outputs/healed/policy_confirm/combo_on25_seed1226/step0050",
+        "num_concurrent": 48,
+        "tokenized_requests": false,
+        "max_retries": 3,
+        "config_source": "/home/henry/Documents/PythonProjects/variable-reap/.venv/lib/python3.12/site-packages/lm_eval/tasks/humaneval/humaneval.yaml"
+      }
+    },
+    "mbpp": {
+      "task": "mbpp",
+      "dataset_path": "google-research-datasets/mbpp",
+      "dataset_name": "full",
+      "test_split": "test",
+      "doc_to_text": "You are an expert Python programmer, and here is your task: {{text}} Your code should pass these tests:\n\n{{test_list[0]}}\n{{test_list[1]}}\n{{test_list[2]}}\n[BEGIN]\n",
+      "doc_to_target": "{% if is_fewshot is defined %}{{code}}\n[DONE]{% else %}{{test_list[0]}}\n{{test_list[1]}}\n{{test_list[2]}}{% endif %}",
+      "unsafe_code": true,
+      "description": "",
+      "target_delimiter": "",
+      "fewshot_delimiter": "\n\n",
+      "fewshot_config": {
+        "sampler": "first_n",
+        "split": null,
+        "process_docs": null,
+        "fewshot_indices": null,
+        "samples": "<function list_fewshot_samples at 0x75dc48d079c0>",
+        "doc_to_text": "You are an expert Python programmer, and here is your task: {{text}} Your code should pass these tests:\n\n{{test_list[0]}}\n{{test_list[1]}}\n{{test_list[2]}}\n[BEGIN]\n",
+        "doc_to_choice": null,
+        "doc_to_target": "{% if is_fewshot is defined %}{{code}}\n[DONE]{% else %}{{test_list[0]}}\n{{test_list[1]}}\n{{test_list[2]}}{% endif %}",
+        "gen_prefix": null,
+        "fewshot_delimiter": "\n\n",
+        "target_delimiter": ""
+      },
+      "num_fewshot": 3,
+      "metric_list": [
+        {
+          "metric": "def pass_at_1(\n    references: Union[str, list[str]], predictions: Union[str, list[list[str]]]\n) -> float:\n    if isinstance(references, str):\n        references = [references]\n    if isinstance(predictions[0], str):\n        predictions = [[p] for p in predictions]\n    return pass_at_k.compute(\n        references=references,\n        predictions=predictions,\n        k=[1],\n    )[0][\"pass@1\"]\n",
+          "aggregation": "mean",
+          "higher_is_better": true
+        }
+      ],
+      "output_type": "generate_until",
+      "generation_kwargs": {
+        "until": [
+          "[DONE]"
+        ],
+        "do_sample": false
+      },
+      "repeats": 1,
+      "should_decontaminate": false,
+      "metadata": {
+        "version": 1.0,
+        "model": "student",
+        "base_url": "http://127.0.0.1:8432/v1/completions",
+        "tokenizer": "outputs/healed/policy_confirm/combo_on25_seed1226/step0050",
+        "num_concurrent": 48,
+        "tokenized_requests": false,
+        "max_retries": 3,
+        "config_source": "/home/henry/Documents/PythonProjects/variable-reap/.venv/lib/python3.12/site-packages/lm_eval/tasks/mbpp/mbpp.yaml"
+      }
+    }
+  },
+  "versions": {
+    "humaneval": 1.0,
+    "mbpp": 1.0
+  },
+  "n-shot": {
+    "humaneval": 0,
+    "mbpp": 3
+  },
+  "higher_is_better": {
+    "humaneval": {
+      "pass_at_k": true,
+      "pass@1": true
+    },
+    "mbpp": {
+      "pass_at_1": true
+    }
+  },
+  "n-samples": {
+    "humaneval": {
+      "original": 164,
+      "effective": 164
+    },
+    "mbpp": {
+      "original": 500,
+      "effective": 500
+    }
+  },
+  "config": {
+    "model": "local-completions",
+    "model_args": {
+      "model": "student",
+      "base_url": "http://127.0.0.1:8432/v1/completions",
+      "tokenizer": "outputs/healed/policy_confirm/combo_on25_seed1226/step0050",
+      "num_concurrent": 48,
+      "tokenized_requests": false,
+      "max_retries": 3
+    },
+    "batch_size": 1,
+    "batch_sizes": [],
+    "device": "cuda:0",
+    "use_cache": null,
+    "limit": null,
+    "bootstrap_iters": 100000,
+    "gen_kwargs": {},
+    "random_seed": 0,
+    "numpy_seed": 1234,
+    "torch_seed": 1234,
+    "fewshot_seed": 1234
+  },
+  "git_hash": "1d6a563",
+  "date": 1784564807.4896429,
+  "pretty_env_info": "PyTorch version: 2.12.1+cu126\nIs debug build: False\nCUDA used to build PyTorch: 12.6\nROCM used to build PyTorch: N/A\n\nOS: Pop!_OS 22.04 LTS (x86_64)\nGCC version: (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0\nClang version: Could not collect\nCMake version: version 3.22.1\nLibc version: glibc-2.35\n\nPython version: 3.12.12 (main, Oct 31 2025, 23:02:31) [Clang 21.1.4 ] (64-bit runtime)\nPython platform: Linux-7.0.11-76070011-generic-x86_64-with-glibc2.35\nIs CUDA available: True\nCUDA runtime version: 12.8.93\nCUDA_MODULE_LOADING set to: \nGPU models and configuration: \nGPU 0: NVIDIA GeForce RTX 3090\nGPU 1: NVIDIA GeForce RTX 3090\nGPU 2: NVIDIA GeForce RTX 3090\n\nNvidia driver version: 580.159.03\ncuDNN version: Could not collect\nIs XPU available: False\nHIP runtime version: N/A\nMIOpen runtime version: N/A\nIs XNNPACK available: True\nCaching allocator config: N/A\n\nCPU:\nArchitecture:                            x86_64\nCPU op-mode(s):                          32-bit, 64-bit\nAddress sizes:                           43 bits physical, 48 bits virtual\nByte Order:                              Little Endian\nCPU(s):                                  48\nOn-line CPU(s) list:                     0-47\nVendor ID:                               AuthenticAMD\nModel name:                              AMD EPYC 7402P 24-Core Processor\nCPU family:                              23\nModel:                                   49\nThread(s) per core:                      2\nCore(s) per socket:                      24\nSocket(s):                               1\nStepping:                                0\nFrequency boost:                         enabled\nCPU max MHz:                             3350.8049\nCPU min MHz:                             1500.0000\nBogoMIPS:                                5600.11\nFlags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sev sev_es\nVirtualization:                          AMD-V\nL1d cache:                               768 KiB (24 instances)\nL1i cache:                               768 KiB (24 instances)\nL2 cache:                                12 MiB (24 instances)\nL3 cache:                                128 MiB (8 instances)\nNUMA node(s):                            1\nNUMA node0 CPU(s):                       0-47\nVulnerability Gather data sampling:      Not affected\nVulnerability Ghostwrite:                Not affected\nVulnerability Indirect target selection: Not affected\nVulnerability Itlb multihit:             Not affected\nVulnerability L1tf:                      Not affected\nVulnerability Mds:                       Not affected\nVulnerability Meltdown:                  Not affected\nVulnerability Mmio stale data:           Not affected\nVulnerability Old microcode:             Not affected\nVulnerability Reg file data sampling:    Not affected\nVulnerability Retbleed:                  Mitigation; untrained return thunk; SMT enabled with STIBP protection\nVulnerability Spec rstack overflow:      Mitigation; Safe RET\nVulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl\nVulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization\nVulnerability Spectre v2:                Mitigation; Retpolines; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected\nVulnerability Srbds:                     Not affected\nVulnerability Tsa:                       Not affected\nVulnerability Tsx async abort:           Not affected\nVulnerability Vmscape:                   Mitigation; IBPB before exit to userspace\n\nVersions of relevant libraries:\n[pip3] Could not collect\n[conda] Could not collect",
+  "transformers_version": "4.57.6",
+  "lm_eval_version": "0.4.12",
+  "upper_git_hash": null,
+  "tokenizer_pad_token": [
+    "<pad>",
+    "50280"
+  ],
+  "tokenizer_eos_token": [
+    "|||IP_ADDRESS|||",
+    "50279"
+  ],
+  "tokenizer_bos_token": [
+    "|||IP_ADDRESS|||",
+    "50279"
+  ],
+  "eot_token_id": 50279,
+  "max_length": 2047,
+  "task_hashes": {
+    "humaneval": "c122632f0bdffdd0162cc8033a879dfe9d3b35e82c52eda6a767069e0a2beb40",
+    "mbpp": "62356f643ed511c895814c093ddb55451941156ba9febb31dc87c6e00d9723e3"
+  },
+  "model_source": "local-completions",
+  "model_name": "student",
+  "model_name_sanitized": "student",
+  "system_instruction": null,
+  "system_instruction_sha": null,
+  "fewshot_as_multiturn": null,
+  "chat_template": null,
+  "chat_template_sha": null,
+  "total_evaluation_time_seconds": "289.5928765599965"
+}

evals/healed_step0112/outputs__healed__keep5_math_run1__step0112/results_2026-07-11T11-00-41.126073.json

@@ -0,0 +1,205 @@
+{
+  "results": {
+    "gsm8k_cot": {
+      "name": "gsm8k_cot",
+      "alias": "gsm8k_cot",
+      "sample_len": 250,
+      "exact_match,strict-match": 0.58,
+      "exact_match_stderr,strict-match": 0.03127799950463661,
+      "exact_match,flexible-extract": 0.584,
+      "exact_match_stderr,flexible-extract": 0.03123585623701455
+    }
+  },
+  "group_subtasks": {},
+  "configs": {
+    "gsm8k_cot": {
+      "task": "gsm8k_cot",
+      "dataset_path": "openai/gsm8k",
+      "dataset_name": "main",
+      "test_split": "test",
+      "doc_to_text": "Q: {{question}}\nA:",
+      "doc_to_target": "{{answer.split('####')[-1].strip() if answer is defined else target}}",
+      "unsafe_code": false,
+      "description": "",
+      "target_delimiter": " ",
+      "fewshot_delimiter": "\n\n",
+      "fewshot_config": {
+        "sampler": "first_n",
+        "split": null,
+        "process_docs": null,
+        "fewshot_indices": null,
+        "samples": [
+          {
+            "question": "There are 15 trees in the grove. Grove workers will plant trees in the grove today. After they are done, there will be 21 trees. How many trees did the grove workers plant today?",
+            "target": "There are 15 trees originally. Then there were 21 trees after some more were planted. So there must have been 21 - 15 = 6. The answer is 6."
+          },
+          {
+            "question": "If there are 3 cars in the parking lot and 2 more cars arrive, how many cars are in the parking lot?",
+            "target": "There are originally 3 cars. 2 more cars arrive. 3 + 2 = 5. The answer is 5."
+          },
+          {
+            "question": "Leah had 32 chocolates and her sister had 42. If they ate 35, how many pieces do they have left in total?",
+            "target": "Originally, Leah had 32 chocolates. Her sister had 42. So in total they had 32 + 42 = 74. After eating 35, they had 74 - 35 = 39. The answer is 39."
+          },
+          {
+            "question": "Jason had 20 lollipops. He gave Denny some lollipops. Now Jason has 12 lollipops. How many lollipops did Jason give to Denny?",
+            "target": "Jason started with 20 lollipops. Then he had 12 after giving some to Denny. So he gave Denny 20 - 12 = 8. The answer is 8."
+          },
+          {
+            "question": "Shawn has five toys. For Christmas, he got two toys each from his mom and dad. How many toys does he have now?",
+            "target": "Shawn started with 5 toys. If he got 2 toys each from his mom and dad, then that is 4 more toys. 5 + 4 = 9. The answer is 9."
+          },
+          {
+            "question": "There were nine computers in the server room. Five more computers were installed each day, from monday to thursday. How many computers are now in the server room?",
+            "target": "There were originally 9 computers. For each of 4 days, 5 more computers were added. So 5 * 4 = 20 computers were added. 9 + 20 is 29. The answer is 29."
+          },
+          {
+            "question": "Michael had 58 golf balls. On tuesday, he lost 23 golf balls. On wednesday, he lost 2 more. How many golf balls did he have at the end of wednesday?",
+            "target": "Michael started with 58 golf balls. After losing 23 on tuesday, he had 58 - 23 = 35. After losing 2 more, he had 35 - 2 = 33 golf balls. The answer is 33."
+          },
+          {
+            "question": "Olivia has $23. She bought five bagels for $3 each. How much money does she have left?",
+            "target": "Olivia had 23 dollars. 5 bagels for 3 dollars each will be 5 x 3 = 15 dollars. So she has 23 - 15 dollars left. 23 - 15 is 8. The answer is 8."
+          }
+        ],
+        "doc_to_text": "Q: {{question}}\nA:",
+        "doc_to_choice": null,
+        "doc_to_target": "{{answer.split('####')[-1].strip() if answer is defined else target}}",
+        "gen_prefix": null,
+        "fewshot_delimiter": "\n\n",
+        "target_delimiter": " "
+      },
+      "num_fewshot": 8,
+      "metric_list": [
+        {
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "ignore_case": true,
+          "ignore_punctuation": false,
+          "metric": "exact_match",
+          "regexes_to_ignore": [
+            ",",
+            "\\$",
+            "(?s).*#### ",
+            "\\.$"
+          ]
+        }
+      ],
+      "output_type": "generate_until",
+      "generation_kwargs": {
+        "do_sample": false,
+        "until": [
+          "Q:",
+          "</s>",
+          "<|im_end|>"
+        ]
+      },
+      "repeats": 1,
+      "filter_list": [
+        {
+          "filter": [
+            {
+              "function": "regex",
+              "regex_pattern": "The answer is (\\-?[0-9\\.\\,]+)."
+            },
+            {
+              "function": "take_first"
+            }
+          ],
+          "name": "strict-match"
+        },
+        {
+          "filter": [
+            {
+              "function": "regex",
+              "group_select": -1,
+              "regex_pattern": "(-?[$0-9.,]{2,})|(-?[0-9]+)"
+            },
+            {
+              "function": "take_first"
+            }
+          ],
+          "name": "flexible-extract"
+        }
+      ],
+      "should_decontaminate": false,
+      "metadata": {
+        "version": 3.0,
+        "pretrained": "outputs/healed/keep5_math_run1/step0112",
+        "dtype": "bfloat16",
+        "trust_remote_code": true,
+        "config_source": "/home/henry/Documents/PythonProjects/variable-reap/.venv/lib/python3.12/site-packages/lm_eval/tasks/gsm8k/gsm8k-cot.yaml"
+      }
+    }
+  },
+  "versions": {
+    "gsm8k_cot": 3.0
+  },
+  "n-shot": {
+    "gsm8k_cot": 8
+  },
+  "higher_is_better": {
+    "gsm8k_cot": {
+      "exact_match": true
+    }
+  },
+  "n-samples": {
+    "gsm8k_cot": {
+      "original": 1319,
+      "effective": 250
+    }
+  },
+  "config": {
+    "model": "hf",
+    "model_args": {
+      "pretrained": "outputs/healed/keep5_math_run1/step0112",
+      "dtype": "bfloat16",
+      "trust_remote_code": true
+    },
+    "model_num_parameters": 3697491968,
+    "model_dtype": "torch.bfloat16",
+    "model_revision": "main",
+    "model_sha": "",
+    "batch_size": "auto",
+    "batch_sizes": [],
+    "device": "cuda:0",
+    "use_cache": null,
+    "limit": 250.0,
+    "bootstrap_iters": 100000,
+    "gen_kwargs": {},
+    "random_seed": 0,
+    "numpy_seed": 1234,
+    "torch_seed": 1234,
+    "fewshot_seed": 1234
+  },
+  "git_hash": "dfc190a",
+  "date": 1783791635.7707028,
+  "pretty_env_info": "'NoneType' object has no attribute 'splitlines'",
+  "transformers_version": "4.57.6",
+  "lm_eval_version": "0.4.12",
+  "upper_git_hash": null,
+  "tokenizer_pad_token": [
+    "<pad>",
+    "50280"
+  ],
+  "tokenizer_eos_token": [
+    "|||IP_ADDRESS|||",
+    "50279"
+  ],
+  "tokenizer_bos_token": [
+    "|||IP_ADDRESS|||",
+    "50279"
+  ],
+  "eot_token_id": 50279,
+  "max_length": 4096,
+  "task_hashes": {},
+  "model_source": "hf",
+  "model_name": "outputs/healed/keep5_math_run1/step0112",
+  "model_name_sanitized": "outputs__healed__keep5_math_run1__step0112",
+  "system_instruction": null,
+  "system_instruction_sha": null,
+  "fewshot_as_multiturn": null,
+  "chat_template": null,
+  "chat_template_sha": null,
+  "total_evaluation_time_seconds": "1211.8154170799971"
+}

evals/healed_step0112_full/outputs__healed__keep5_math_run1__step0112/results_2026-07-11T12-36-52.361127.json

@@ -0,0 +1,205 @@
+{
+  "results": {
+    "gsm8k_cot": {
+      "name": "gsm8k_cot",
+      "alias": "gsm8k_cot",
+      "sample_len": 1319,
+      "exact_match,strict-match": 0.5678544351781653,
+      "exact_match_stderr,strict-match": 0.013645072137842442,
+      "exact_match,flexible-extract": 0.5708870356330553,
+      "exact_match_stderr,flexible-extract": 0.013633369425647232
+    }
+  },
+  "group_subtasks": {},
+  "configs": {
+    "gsm8k_cot": {
+      "task": "gsm8k_cot",
+      "dataset_path": "openai/gsm8k",
+      "dataset_name": "main",
+      "test_split": "test",
+      "doc_to_text": "Q: {{question}}\nA:",
+      "doc_to_target": "{{answer.split('####')[-1].strip() if answer is defined else target}}",
+      "unsafe_code": false,
+      "description": "",
+      "target_delimiter": " ",
+      "fewshot_delimiter": "\n\n",
+      "fewshot_config": {
+        "sampler": "first_n",
+        "split": null,
+        "process_docs": null,
+        "fewshot_indices": null,
+        "samples": [
+          {
+            "question": "There are 15 trees in the grove. Grove workers will plant trees in the grove today. After they are done, there will be 21 trees. How many trees did the grove workers plant today?",
+            "target": "There are 15 trees originally. Then there were 21 trees after some more were planted. So there must have been 21 - 15 = 6. The answer is 6."
+          },
+          {
+            "question": "If there are 3 cars in the parking lot and 2 more cars arrive, how many cars are in the parking lot?",
+            "target": "There are originally 3 cars. 2 more cars arrive. 3 + 2 = 5. The answer is 5."
+          },
+          {
+            "question": "Leah had 32 chocolates and her sister had 42. If they ate 35, how many pieces do they have left in total?",
+            "target": "Originally, Leah had 32 chocolates. Her sister had 42. So in total they had 32 + 42 = 74. After eating 35, they had 74 - 35 = 39. The answer is 39."
+          },
+          {
+            "question": "Jason had 20 lollipops. He gave Denny some lollipops. Now Jason has 12 lollipops. How many lollipops did Jason give to Denny?",
+            "target": "Jason started with 20 lollipops. Then he had 12 after giving some to Denny. So he gave Denny 20 - 12 = 8. The answer is 8."
+          },
+          {
+            "question": "Shawn has five toys. For Christmas, he got two toys each from his mom and dad. How many toys does he have now?",
+            "target": "Shawn started with 5 toys. If he got 2 toys each from his mom and dad, then that is 4 more toys. 5 + 4 = 9. The answer is 9."
+          },
+          {
+            "question": "There were nine computers in the server room. Five more computers were installed each day, from monday to thursday. How many computers are now in the server room?",
+            "target": "There were originally 9 computers. For each of 4 days, 5 more computers were added. So 5 * 4 = 20 computers were added. 9 + 20 is 29. The answer is 29."
+          },
+          {
+            "question": "Michael had 58 golf balls. On tuesday, he lost 23 golf balls. On wednesday, he lost 2 more. How many golf balls did he have at the end of wednesday?",
+            "target": "Michael started with 58 golf balls. After losing 23 on tuesday, he had 58 - 23 = 35. After losing 2 more, he had 35 - 2 = 33 golf balls. The answer is 33."
+          },
+          {
+            "question": "Olivia has $23. She bought five bagels for $3 each. How much money does she have left?",
+            "target": "Olivia had 23 dollars. 5 bagels for 3 dollars each will be 5 x 3 = 15 dollars. So she has 23 - 15 dollars left. 23 - 15 is 8. The answer is 8."
+          }
+        ],
+        "doc_to_text": "Q: {{question}}\nA:",
+        "doc_to_choice": null,
+        "doc_to_target": "{{answer.split('####')[-1].strip() if answer is defined else target}}",
+        "gen_prefix": null,
+        "fewshot_delimiter": "\n\n",
+        "target_delimiter": " "
+      },
+      "num_fewshot": 8,
+      "metric_list": [
+        {
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "ignore_case": true,
+          "ignore_punctuation": false,
+          "metric": "exact_match",
+          "regexes_to_ignore": [
+            ",",
+            "\\$",
+            "(?s).*#### ",
+            "\\.$"
+          ]
+        }
+      ],
+      "output_type": "generate_until",
+      "generation_kwargs": {
+        "do_sample": false,
+        "until": [
+          "Q:",
+          "</s>",
+          "<|im_end|>"
+        ]
+      },
+      "repeats": 1,
+      "filter_list": [
+        {
+          "filter": [
+            {
+              "function": "regex",
+              "regex_pattern": "The answer is (\\-?[0-9\\.\\,]+)."
+            },
+            {
+              "function": "take_first"
+            }
+          ],
+          "name": "strict-match"
+        },
+        {
+          "filter": [
+            {
+              "function": "regex",
+              "group_select": -1,
+              "regex_pattern": "(-?[$0-9.,]{2,})|(-?[0-9]+)"
+            },
+            {
+              "function": "take_first"
+            }
+          ],
+          "name": "flexible-extract"
+        }
+      ],
+      "should_decontaminate": false,
+      "metadata": {
+        "version": 3.0,
+        "pretrained": "outputs/healed/keep5_math_run1/step0112",
+        "dtype": "bfloat16",
+        "trust_remote_code": true,
+        "config_source": "/home/henry/Documents/PythonProjects/variable-reap/.venv/lib/python3.12/site-packages/lm_eval/tasks/gsm8k/gsm8k-cot.yaml"
+      }
+    }
+  },
+  "versions": {
+    "gsm8k_cot": 3.0
+  },
+  "n-shot": {
+    "gsm8k_cot": 8
+  },
+  "higher_is_better": {
+    "gsm8k_cot": {
+      "exact_match": true
+    }
+  },
+  "n-samples": {
+    "gsm8k_cot": {
+      "original": 1319,
+      "effective": 1319
+    }
+  },
+  "config": {
+    "model": "hf",
+    "model_args": {
+      "pretrained": "outputs/healed/keep5_math_run1/step0112",
+      "dtype": "bfloat16",
+      "trust_remote_code": true
+    },
+    "model_num_parameters": 3697491968,
+    "model_dtype": "torch.bfloat16",
+    "model_revision": "main",
+    "model_sha": "",
+    "batch_size": "auto",
+    "batch_sizes": [],
+    "device": "cuda:0",
+    "use_cache": null,
+    "limit": null,
+    "bootstrap_iters": 100000,
+    "gen_kwargs": {},
+    "random_seed": 0,
+    "numpy_seed": 1234,
+    "torch_seed": 1234,
+    "fewshot_seed": 1234
+  },
+  "git_hash": "59171ce",
+  "date": 1783792886.9500268,
+  "pretty_env_info": "'NoneType' object has no attribute 'splitlines'",
+  "transformers_version": "4.57.6",
+  "lm_eval_version": "0.4.12",
+  "upper_git_hash": null,
+  "tokenizer_pad_token": [
+    "<pad>",
+    "50280"
+  ],
+  "tokenizer_eos_token": [
+    "|||IP_ADDRESS|||",
+    "50279"
+  ],
+  "tokenizer_bos_token": [
+    "|||IP_ADDRESS|||",
+    "50279"
+  ],
+  "eot_token_id": 50279,
+  "max_length": 4096,
+  "task_hashes": {},
+  "model_source": "hf",
+  "model_name": "outputs/healed/keep5_math_run1/step0112",
+  "model_name_sanitized": "outputs__healed__keep5_math_run1__step0112",
+  "system_instruction": null,
+  "system_instruction_sha": null,
+  "fewshot_as_multiturn": null,
+  "chat_template": null,
+  "chat_template_sha": null,
+  "total_evaluation_time_seconds": "5731.806047619"
+}

evals/healing_breadth/keep25_traj/step0025_chat.json.server.log

@@ -0,0 +1,138 @@
+(APIServer pid=311884) INFO 07-15 19:41:01 [api_utils.py:339] 
+(APIServer pid=311884) INFO 07-15 19:41:01 [api_utils.py:339]        █     █     █▄   ▄█
+(APIServer pid=311884) INFO 07-15 19:41:01 [api_utils.py:339]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.25.0
+(APIServer pid=311884) INFO 07-15 19:41:01 [api_utils.py:339]   █▄█▀ █     █     █     █  model   outputs/healed/healing_breadth/glean_math_keep25_seed1224/step0025
+(APIServer pid=311884) INFO 07-15 19:41:01 [api_utils.py:339]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
+(APIServer pid=311884) INFO 07-15 19:41:01 [api_utils.py:339] 
+(APIServer pid=311884) INFO 07-15 19:41:01 [api_utils.py:273] non-default args: {'model_tag': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224/step0025', 'host': '127.0.0.1', 'port': 8377, 'model': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224/step0025', 'max_model_len': 2048, 'enforce_eager': True, 'served_model_name': ['student'], 'gpu_memory_utilization': 0.85}
+(APIServer pid=311884) INFO 07-15 19:41:01 [model.py:619] Resolved architecture: PrunedOlmoeForCausalLM
+(APIServer pid=311884) INFO 07-15 19:41:01 [model.py:1776] Using max model len 2048
+(APIServer pid=311884) INFO 07-15 19:41:01 [vllm.py:1042] Asynchronous scheduling is enabled.
+(APIServer pid=311884) WARNING 07-15 19:41:01 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(APIServer pid=311884) WARNING 07-15 19:41:01 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(APIServer pid=311884) INFO 07-15 19:41:01 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(APIServer pid=311884) INFO 07-15 19:41:01 [vllm.py:1322] Cudagraph is disabled under eager mode
+(APIServer pid=311884) INFO 07-15 19:41:01 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(EngineCore pid=312001) INFO 07-15 19:41:08 [core.py:114] Initializing a V1 LLM engine (v0.25.0) with config: model='outputs/healed/healing_breadth/glean_math_keep25_seed1224/step0025', speculative_config=None, tokenizer='outputs/healed/healing_breadth/glean_math_keep25_seed1224/step0025', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=student, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.NONE: 0>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, moe_backend='auto', linear_backend='auto')
+(EngineCore pid=312001) INFO 07-15 19:41:09 [parallel_state.py:1607] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.0.15:40487 backend=nccl
+(EngineCore pid=312001) INFO 07-15 19:41:09 [parallel_state.py:1942] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A
+(EngineCore pid=312001) INFO 07-15 19:41:10 [topk_topp_sampler.py:55] Using FlashInfer for top-p & top-k sampling.
+(EngineCore pid=312001) INFO 07-15 19:41:10 [gpu_model_runner.py:5209] Starting to load model outputs/healed/healing_breadth/glean_math_keep25_seed1224/step0025...
+(EngineCore pid=312001) INFO 07-15 19:41:10 [cuda.py:476] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
+(EngineCore pid=312001) INFO 07-15 19:41:10 [flash_attn.py:718] Using FlashAttention version 2
+(EngineCore pid=312001) /home/henry/.cache/glean/megablocks-variable-93a1479bc15b/megablocks/grouped_gemm_util.py:10: UserWarning: Grouped GEMM not available.
+(EngineCore pid=312001)   warnings.warn('Grouped GEMM not available.')
+(EngineCore pid=312001) INFO 07-15 19:41:10 [weight_utils.py:849] Filesystem type for checkpoints: EXT4. Checkpoint size: 3.89 GiB. Available RAM: 109.18 GiB.
+(EngineCore pid=312001) INFO 07-15 19:41:10 [weight_utils.py:872] Auto-prefetch is disabled because the filesystem (EXT4) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.
+(EngineCore pid=312001) 
+(EngineCore pid=312001) 
+(EngineCore pid=312001) 
+(EngineCore pid=312001) 
+(EngineCore pid=312001) INFO 07-15 19:41:16 [default_loader.py:430] Loading weights took 5.97 seconds
+(EngineCore pid=312001) INFO 07-15 19:41:17 [gpu_model_runner.py:5306] Model loading took 3.89 GiB memory and 6.151022 seconds
+(EngineCore pid=312001) INFO 07-15 19:41:18 [gpu_worker.py:538] Available KV cache memory: 15.82 GiB
+(EngineCore pid=312001) INFO 07-15 19:41:18 [kv_cache_utils.py:2146] GPU KV cache size: 129,584 tokens
+(EngineCore pid=312001) INFO 07-15 19:41:18 [kv_cache_utils.py:2147] Maximum concurrency for 2,048 tokens per request: 63.27x
+(EngineCore pid=312001) INFO 07-15 19:41:19 [cutedsl_warmup.py:97] Skipping CuTeDSL warmup because no compile units were requested.
+(EngineCore pid=312001) INFO 07-15 19:41:19 [jit_monitor.py:73] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
+(EngineCore pid=312001) INFO 07-15 19:41:19 [core.py:344] init engine (profile, create kv cache, warmup model) took 2.06 s
+(EngineCore pid=312001) INFO 07-15 19:41:19 [vllm.py:1042] Asynchronous scheduling is enabled.
+(EngineCore pid=312001) WARNING 07-15 19:41:19 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(EngineCore pid=312001) WARNING 07-15 19:41:19 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(EngineCore pid=312001) INFO 07-15 19:41:19 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(EngineCore pid=312001) INFO 07-15 19:41:19 [vllm.py:1322] Cudagraph is disabled under eager mode
+(EngineCore pid=312001) INFO 07-15 19:41:19 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(APIServer pid=311884) INFO 07-15 19:41:19 [api_server.py:612] Supported tasks: ['generate']
+(APIServer pid=311884) WARNING 07-15 19:41:19 [__init__.py:36] SECURITY WARNING: Development endpoints are enabled! This should NOT be used in production!
+(APIServer pid=311884) INFO 07-15 19:41:19 [hf.py:548] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
+(APIServer pid=311884) INFO 07-15 19:41:19 [api_server.py:616] Starting vLLM server on http://127.0.0.1:8377
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:37] Available routes are:
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /openapi.json, Methods: GET, HEAD
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /docs, Methods: GET, HEAD
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: GET, HEAD
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /redoc, Methods: GET, HEAD
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /load, Methods: GET
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /version, Methods: GET
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /health, Methods: GET
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /metrics, Methods: GET
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /tokenize, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /detokenize, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /v1/models, Methods: GET
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /ping, Methods: GET
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /ping, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /invocations, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /reset_prefix_cache, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /reset_mm_cache, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /reset_encoder_cache, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /pause, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /resume, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /is_paused, Methods: GET
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /init_weight_transfer_engine, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /start_weight_update, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /update_weights, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /finish_weight_update, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /get_world_size, Methods: GET
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /collective_rpc, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /server_info, Methods: GET
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /sleep, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /wake_up, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /is_sleeping, Methods: GET
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /v1/chat/completions/batch, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /v1/responses, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /v1/completions, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /v1/messages, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /generative_scoring, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /v1/completions/render, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /v1/chat/completions/derender, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /v1/completions/derender, Methods: POST
+(APIServer pid=311884) INFO 07-15 19:41:19 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
+(APIServer pid=311884) INFO:     Started server process [311884]
+(APIServer pid=311884) INFO:     Waiting for application startup.
+(APIServer pid=311884) INFO:     Application startup complete.
+(APIServer pid=311884) INFO:     127.0.0.1:48568 - "GET /health HTTP/1.1" 200 OK
+(EngineCore pid=312001) WARNING 07-15 19:41:22 [jit_monitor.py:129] Triton kernel JIT compilation during inference: _build_route_rows. This causes a latency spike; consider extending warmup to cover this shape/config.
+(APIServer pid=311884) INFO 07-15 19:41:30 [loggers.py:273] Engine 000: Avg prompt throughput: 2053.5 tokens/s, Avg generation throughput: 2071.8 tokens/s, Running: 256 reqs, Waiting: 1058 reqs, GPU KV cache usage: 34.5%, Prefix cache hit rate: 93.1%
+(APIServer pid=311884) INFO 07-15 19:41:40 [loggers.py:273] Engine 000: Avg prompt throughput: 218.5 tokens/s, Avg generation throughput: 3017.7 tokens/s, Running: 256 reqs, Waiting: 1029 reqs, GPU KV cache usage: 54.1%, Prefix cache hit rate: 93.1%
+(APIServer pid=311884) INFO 07-15 19:41:50 [loggers.py:273] Engine 000: Avg prompt throughput: 480.7 tokens/s, Avg generation throughput: 2885.8 tokens/s, Running: 256 reqs, Waiting: 968 reqs, GPU KV cache usage: 64.9%, Prefix cache hit rate: 93.2%
+(APIServer pid=311884) INFO 07-15 19:42:00 [loggers.py:273] Engine 000: Avg prompt throughput: 640.2 tokens/s, Avg generation throughput: 2833.2 tokens/s, Running: 256 reqs, Waiting: 889 reqs, GPU KV cache usage: 66.8%, Prefix cache hit rate: 93.2%
+(APIServer pid=311884) INFO 07-15 19:42:10 [loggers.py:273] Engine 000: Avg prompt throughput: 956.4 tokens/s, Avg generation throughput: 2776.9 tokens/s, Running: 256 reqs, Waiting: 763 reqs, GPU KV cache usage: 43.7%, Prefix cache hit rate: 93.3%
+(APIServer pid=311884) INFO 07-15 19:42:20 [loggers.py:273] Engine 000: Avg prompt throughput: 484.7 tokens/s, Avg generation throughput: 2962.5 tokens/s, Running: 256 reqs, Waiting: 701 reqs, GPU KV cache usage: 52.9%, Prefix cache hit rate: 93.3%
+(APIServer pid=311884) INFO 07-15 19:42:30 [loggers.py:273] Engine 000: Avg prompt throughput: 614.9 tokens/s, Avg generation throughput: 2910.0 tokens/s, Running: 256 reqs, Waiting: 622 reqs, GPU KV cache usage: 57.1%, Prefix cache hit rate: 93.3%
+(APIServer pid=311884) INFO 07-15 19:42:40 [loggers.py:273] Engine 000: Avg prompt throughput: 672.6 tokens/s, Avg generation throughput: 2884.2 tokens/s, Running: 255 reqs, Waiting: 536 reqs, GPU KV cache usage: 57.4%, Prefix cache hit rate: 93.3%
+(APIServer pid=311884) INFO 07-15 19:42:50 [loggers.py:273] Engine 000: Avg prompt throughput: 670.0 tokens/s, Avg generation throughput: 2858.7 tokens/s, Running: 256 reqs, Waiting: 452 reqs, GPU KV cache usage: 57.2%, Prefix cache hit rate: 93.3%
+(APIServer pid=311884) INFO 07-15 19:43:00 [loggers.py:273] Engine 000: Avg prompt throughput: 664.9 tokens/s, Avg generation throughput: 2910.0 tokens/s, Running: 256 reqs, Waiting: 369 reqs, GPU KV cache usage: 56.2%, Prefix cache hit rate: 93.4%
+(APIServer pid=311884) INFO 07-15 19:43:10 [loggers.py:273] Engine 000: Avg prompt throughput: 680.4 tokens/s, Avg generation throughput: 2884.4 tokens/s, Running: 252 reqs, Waiting: 289 reqs, GPU KV cache usage: 58.2%, Prefix cache hit rate: 93.4%
+(APIServer pid=311884) INFO 07-15 19:43:20 [loggers.py:273] Engine 000: Avg prompt throughput: 746.7 tokens/s, Avg generation throughput: 2857.6 tokens/s, Running: 256 reqs, Waiting: 194 reqs, GPU KV cache usage: 55.9%, Prefix cache hit rate: 93.4%
+(APIServer pid=311884) INFO 07-15 19:43:30 [loggers.py:273] Engine 000: Avg prompt throughput: 514.6 tokens/s, Avg generation throughput: 2886.1 tokens/s, Running: 254 reqs, Waiting: 127 reqs, GPU KV cache usage: 58.7%, Prefix cache hit rate: 93.4%
+(APIServer pid=311884) INFO 07-15 19:43:40 [loggers.py:273] Engine 000: Avg prompt throughput: 703.3 tokens/s, Avg generation throughput: 2884.4 tokens/s, Running: 255 reqs, Waiting: 44 reqs, GPU KV cache usage: 58.5%, Prefix cache hit rate: 93.4%
+(APIServer pid=311884) INFO 07-15 19:43:50 [loggers.py:273] Engine 000: Avg prompt throughput: 344.7 tokens/s, Avg generation throughput: 2844.3 tokens/s, Running: 216 reqs, Waiting: 0 reqs, GPU KV cache usage: 54.5%, Prefix cache hit rate: 93.4%
+(APIServer pid=311884) INFO 07-15 19:44:00 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2489.6 tokens/s, Running: 104 reqs, Waiting: 0 reqs, GPU KV cache usage: 33.7%, Prefix cache hit rate: 93.4%
+(APIServer pid=311884) INFO 07-15 19:44:10 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1321.5 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.1%, Prefix cache hit rate: 93.4%
+(APIServer pid=311884) INFO:     127.0.0.1:48570 - "POST /v1/completions HTTP/1.1" 200 OK
+(EngineCore pid=312001) INFO 07-15 19:44:11 [core.py:1214] [shutdown] EngineCore: trigger received signal=SIGTERM
+(APIServer pid=311884) INFO 07-15 19:44:11 [launcher.py:100] [shutdown] API server: shutdown triggered
+(APIServer pid=311884) INFO 07-15 19:44:11 [launcher.py:116] [shutdown] API server: stopping engine client mode=abort timeout=0s
+(EngineCore pid=312001) INFO 07-15 19:44:11 [core.py:1333] [shutdown] EngineCore: start mode=abort timeout=0s
+(EngineCore pid=312001) INFO 07-15 19:44:11 [core.py:1364] [shutdown] EngineCore: request processing complete; starting resource teardown
+(EngineCore pid=312001) INFO 07-15 19:44:11 [core.py:1227] [shutdown] EngineCore: exiting busy loop
+(APIServer pid=311884) INFO 07-15 19:44:11 [core_client.py:655] [shutdown] MPClient: start timeout=0s
+(APIServer pid=311884) INFO 07-15 19:44:11 [core_client.py:657] [shutdown] MPClient: stopping engine manager
+(APIServer pid=311884) WARNING 07-15 19:44:11 [utils.py:626] [shutdown] Process manager: force killing remaining processes count=1
+(APIServer pid=311884) INFO 07-15 19:44:11 [core_client.py:659] [shutdown] MPClient: engine manager stopped
+(APIServer pid=311884) INFO 07-15 19:44:11 [core_client.py:660] [shutdown] MPClient: cleaning up background resources
+(APIServer pid=311884) INFO 07-15 19:44:11 [core_client.py:662] [shutdown] MPClient: complete
+(APIServer pid=311884) INFO 07-15 19:44:11 [launcher.py:125] [shutdown] API server: engine client stopped
+(APIServer pid=311884) INFO 07-15 19:44:11 [launcher.py:128] [shutdown] API server: signalling HTTP server shutdown
+(APIServer pid=311884) INFO 07-15 19:44:11 [launcher.py:149] [shutdown] API server: shutting down FastAPI HTTP server
+(APIServer pid=311884) INFO:     Shutting down
+(APIServer pid=311884) INFO:     Waiting for application shutdown.
+(APIServer pid=311884) INFO:     Application shutdown complete.
+/home/henry/.local/share/uv/python/cpython-3.12.12-linux-x86_64-gnu/lib/python3.12/multiprocessing/resource_tracker.py:279: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
+  warnings.warn('resource_tracker: There appear to be %d '

evals/healing_breadth/keep25_traj/step0050_chat.json.server.log

@@ -0,0 +1,138 @@
+(APIServer pid=312789) INFO 07-15 19:44:36 [api_utils.py:339] 
+(APIServer pid=312789) INFO 07-15 19:44:36 [api_utils.py:339]        █     █     █▄   ▄█
+(APIServer pid=312789) INFO 07-15 19:44:36 [api_utils.py:339]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.25.0
+(APIServer pid=312789) INFO 07-15 19:44:36 [api_utils.py:339]   █▄█▀ █     █     █     █  model   outputs/healed/healing_breadth/glean_math_keep25_seed1224/step0050
+(APIServer pid=312789) INFO 07-15 19:44:36 [api_utils.py:339]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
+(APIServer pid=312789) INFO 07-15 19:44:36 [api_utils.py:339] 
+(APIServer pid=312789) INFO 07-15 19:44:36 [api_utils.py:273] non-default args: {'model_tag': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224/step0050', 'host': '127.0.0.1', 'port': 8377, 'model': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224/step0050', 'max_model_len': 2048, 'enforce_eager': True, 'served_model_name': ['student'], 'gpu_memory_utilization': 0.85}
+(APIServer pid=312789) INFO 07-15 19:44:36 [model.py:619] Resolved architecture: PrunedOlmoeForCausalLM
+(APIServer pid=312789) INFO 07-15 19:44:36 [model.py:1776] Using max model len 2048
+(APIServer pid=312789) INFO 07-15 19:44:36 [vllm.py:1042] Asynchronous scheduling is enabled.
+(APIServer pid=312789) WARNING 07-15 19:44:36 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(APIServer pid=312789) WARNING 07-15 19:44:36 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(APIServer pid=312789) INFO 07-15 19:44:36 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(APIServer pid=312789) INFO 07-15 19:44:36 [vllm.py:1322] Cudagraph is disabled under eager mode
+(APIServer pid=312789) INFO 07-15 19:44:36 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(EngineCore pid=312907) INFO 07-15 19:44:43 [core.py:114] Initializing a V1 LLM engine (v0.25.0) with config: model='outputs/healed/healing_breadth/glean_math_keep25_seed1224/step0050', speculative_config=None, tokenizer='outputs/healed/healing_breadth/glean_math_keep25_seed1224/step0050', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=student, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.NONE: 0>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, moe_backend='auto', linear_backend='auto')
+(EngineCore pid=312907) INFO 07-15 19:44:44 [parallel_state.py:1607] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.0.15:38777 backend=nccl
+(EngineCore pid=312907) INFO 07-15 19:44:44 [parallel_state.py:1942] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A
+(EngineCore pid=312907) INFO 07-15 19:44:44 [topk_topp_sampler.py:55] Using FlashInfer for top-p & top-k sampling.
+(EngineCore pid=312907) INFO 07-15 19:44:44 [gpu_model_runner.py:5209] Starting to load model outputs/healed/healing_breadth/glean_math_keep25_seed1224/step0050...
+(EngineCore pid=312907) INFO 07-15 19:44:45 [cuda.py:476] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
+(EngineCore pid=312907) INFO 07-15 19:44:45 [flash_attn.py:718] Using FlashAttention version 2
+(EngineCore pid=312907) /home/henry/.cache/glean/megablocks-variable-93a1479bc15b/megablocks/grouped_gemm_util.py:10: UserWarning: Grouped GEMM not available.
+(EngineCore pid=312907)   warnings.warn('Grouped GEMM not available.')
+(EngineCore pid=312907) INFO 07-15 19:44:45 [weight_utils.py:849] Filesystem type for checkpoints: EXT4. Checkpoint size: 3.89 GiB. Available RAM: 109.18 GiB.
+(EngineCore pid=312907) INFO 07-15 19:44:45 [weight_utils.py:872] Auto-prefetch is disabled because the filesystem (EXT4) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.
+(EngineCore pid=312907) 
+(EngineCore pid=312907) 
+(EngineCore pid=312907) 
+(EngineCore pid=312907) 
+(EngineCore pid=312907) INFO 07-15 19:44:47 [default_loader.py:430] Loading weights took 2.51 seconds
+(EngineCore pid=312907) INFO 07-15 19:44:48 [gpu_model_runner.py:5306] Model loading took 3.89 GiB memory and 2.695594 seconds
+(EngineCore pid=312907) INFO 07-15 19:44:50 [gpu_worker.py:538] Available KV cache memory: 15.82 GiB
+(EngineCore pid=312907) INFO 07-15 19:44:50 [kv_cache_utils.py:2146] GPU KV cache size: 129,584 tokens
+(EngineCore pid=312907) INFO 07-15 19:44:50 [kv_cache_utils.py:2147] Maximum concurrency for 2,048 tokens per request: 63.27x
+(EngineCore pid=312907) INFO 07-15 19:44:50 [cutedsl_warmup.py:97] Skipping CuTeDSL warmup because no compile units were requested.
+(EngineCore pid=312907) INFO 07-15 19:44:50 [jit_monitor.py:73] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
+(EngineCore pid=312907) INFO 07-15 19:44:50 [core.py:344] init engine (profile, create kv cache, warmup model) took 2.08 s
+(EngineCore pid=312907) INFO 07-15 19:44:50 [vllm.py:1042] Asynchronous scheduling is enabled.
+(EngineCore pid=312907) WARNING 07-15 19:44:50 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(EngineCore pid=312907) WARNING 07-15 19:44:50 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(EngineCore pid=312907) INFO 07-15 19:44:50 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(EngineCore pid=312907) INFO 07-15 19:44:50 [vllm.py:1322] Cudagraph is disabled under eager mode
+(EngineCore pid=312907) INFO 07-15 19:44:50 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(APIServer pid=312789) INFO 07-15 19:44:50 [api_server.py:612] Supported tasks: ['generate']
+(APIServer pid=312789) WARNING 07-15 19:44:50 [__init__.py:36] SECURITY WARNING: Development endpoints are enabled! This should NOT be used in production!
+(APIServer pid=312789) INFO 07-15 19:44:50 [hf.py:548] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
+(APIServer pid=312789) INFO 07-15 19:44:50 [api_server.py:616] Starting vLLM server on http://127.0.0.1:8377
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:37] Available routes are:
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /openapi.json, Methods: GET, HEAD
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /docs, Methods: GET, HEAD
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: GET, HEAD
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /redoc, Methods: GET, HEAD
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /load, Methods: GET
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /version, Methods: GET
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /health, Methods: GET
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /metrics, Methods: GET
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /tokenize, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /detokenize, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /v1/models, Methods: GET
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /ping, Methods: GET
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /ping, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /invocations, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /reset_prefix_cache, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /reset_mm_cache, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /reset_encoder_cache, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /pause, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /resume, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /is_paused, Methods: GET
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /init_weight_transfer_engine, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /start_weight_update, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /update_weights, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /finish_weight_update, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /get_world_size, Methods: GET
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /collective_rpc, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /server_info, Methods: GET
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /sleep, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /wake_up, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /is_sleeping, Methods: GET
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /v1/chat/completions/batch, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /v1/responses, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /v1/completions, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /v1/messages, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /generative_scoring, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /v1/completions/render, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /v1/chat/completions/derender, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /v1/completions/derender, Methods: POST
+(APIServer pid=312789) INFO 07-15 19:44:50 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
+(APIServer pid=312789) INFO:     Started server process [312789]
+(APIServer pid=312789) INFO:     Waiting for application startup.
+(APIServer pid=312789) INFO:     Application startup complete.
+(APIServer pid=312789) INFO:     127.0.0.1:52162 - "GET /health HTTP/1.1" 200 OK
+(EngineCore pid=312907) WARNING 07-15 19:44:52 [jit_monitor.py:129] Triton kernel JIT compilation during inference: _build_route_rows. This causes a latency spike; consider extending warmup to cover this shape/config.
+(APIServer pid=312789) INFO 07-15 19:45:01 [loggers.py:273] Engine 000: Avg prompt throughput: 2049.4 tokens/s, Avg generation throughput: 2297.2 tokens/s, Running: 256 reqs, Waiting: 1059 reqs, GPU KV cache usage: 36.4%, Prefix cache hit rate: 93.1%
+(APIServer pid=312789) INFO 07-15 19:45:11 [loggers.py:273] Engine 000: Avg prompt throughput: 334.0 tokens/s, Avg generation throughput: 3015.6 tokens/s, Running: 256 reqs, Waiting: 1016 reqs, GPU KV cache usage: 54.6%, Prefix cache hit rate: 93.1%
+(APIServer pid=312789) INFO 07-15 19:45:21 [loggers.py:273] Engine 000: Avg prompt throughput: 449.3 tokens/s, Avg generation throughput: 2887.1 tokens/s, Running: 255 reqs, Waiting: 959 reqs, GPU KV cache usage: 65.5%, Prefix cache hit rate: 93.2%
+(APIServer pid=312789) INFO 07-15 19:45:31 [loggers.py:273] Engine 000: Avg prompt throughput: 330.4 tokens/s, Avg generation throughput: 2836.6 tokens/s, Running: 255 reqs, Waiting: 914 reqs, GPU KV cache usage: 77.2%, Prefix cache hit rate: 93.3%
+(APIServer pid=312789) INFO 07-15 19:45:41 [loggers.py:273] Engine 000: Avg prompt throughput: 1164.4 tokens/s, Avg generation throughput: 2698.7 tokens/s, Running: 256 reqs, Waiting: 767 reqs, GPU KV cache usage: 44.4%, Prefix cache hit rate: 93.3%
+(APIServer pid=312789) INFO 07-15 19:45:51 [loggers.py:273] Engine 000: Avg prompt throughput: 367.5 tokens/s, Avg generation throughput: 2964.2 tokens/s, Running: 255 reqs, Waiting: 717 reqs, GPU KV cache usage: 56.0%, Prefix cache hit rate: 93.3%
+(APIServer pid=312789) INFO 07-15 19:46:01 [loggers.py:273] Engine 000: Avg prompt throughput: 616.8 tokens/s, Avg generation throughput: 2884.5 tokens/s, Running: 256 reqs, Waiting: 639 reqs, GPU KV cache usage: 58.4%, Prefix cache hit rate: 93.3%
+(APIServer pid=312789) INFO 07-15 19:46:11 [loggers.py:273] Engine 000: Avg prompt throughput: 560.6 tokens/s, Avg generation throughput: 2859.5 tokens/s, Running: 256 reqs, Waiting: 568 reqs, GPU KV cache usage: 61.9%, Prefix cache hit rate: 93.3%
+(APIServer pid=312789) INFO 07-15 19:46:21 [loggers.py:273] Engine 000: Avg prompt throughput: 612.2 tokens/s, Avg generation throughput: 2833.5 tokens/s, Running: 254 reqs, Waiting: 492 reqs, GPU KV cache usage: 64.3%, Prefix cache hit rate: 93.3%
+(APIServer pid=312789) INFO 07-15 19:46:31 [loggers.py:273] Engine 000: Avg prompt throughput: 782.3 tokens/s, Avg generation throughput: 2882.7 tokens/s, Running: 254 reqs, Waiting: 391 reqs, GPU KV cache usage: 55.2%, Prefix cache hit rate: 93.4%
+(APIServer pid=312789) INFO 07-15 19:46:41 [loggers.py:273] Engine 000: Avg prompt throughput: 610.3 tokens/s, Avg generation throughput: 2885.4 tokens/s, Running: 256 reqs, Waiting: 319 reqs, GPU KV cache usage: 60.4%, Prefix cache hit rate: 93.3%
+(APIServer pid=312789) INFO 07-15 19:46:51 [loggers.py:273] Engine 000: Avg prompt throughput: 544.4 tokens/s, Avg generation throughput: 2885.3 tokens/s, Running: 256 reqs, Waiting: 252 reqs, GPU KV cache usage: 63.7%, Prefix cache hit rate: 93.4%
+(APIServer pid=312789) INFO 07-15 19:47:01 [loggers.py:273] Engine 000: Avg prompt throughput: 543.9 tokens/s, Avg generation throughput: 2860.5 tokens/s, Running: 256 reqs, Waiting: 184 reqs, GPU KV cache usage: 65.5%, Prefix cache hit rate: 93.4%
+(APIServer pid=312789) INFO 07-15 19:47:11 [loggers.py:273] Engine 000: Avg prompt throughput: 753.8 tokens/s, Avg generation throughput: 2831.5 tokens/s, Running: 256 reqs, Waiting: 90 reqs, GPU KV cache usage: 57.4%, Prefix cache hit rate: 93.4%
+(APIServer pid=312789) INFO 07-15 19:47:21 [loggers.py:273] Engine 000: Avg prompt throughput: 654.4 tokens/s, Avg generation throughput: 2883.9 tokens/s, Running: 256 reqs, Waiting: 10 reqs, GPU KV cache usage: 56.6%, Prefix cache hit rate: 93.4%
+(APIServer pid=312789) INFO 07-15 19:47:31 [loggers.py:273] Engine 000: Avg prompt throughput: 74.4 tokens/s, Avg generation throughput: 2786.2 tokens/s, Running: 193 reqs, Waiting: 0 reqs, GPU KV cache usage: 54.0%, Prefix cache hit rate: 93.4%
+(APIServer pid=312789) INFO 07-15 19:47:41 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2309.1 tokens/s, Running: 94 reqs, Waiting: 0 reqs, GPU KV cache usage: 35.4%, Prefix cache hit rate: 93.4%
+(APIServer pid=312789) INFO:     127.0.0.1:52178 - "POST /v1/completions HTTP/1.1" 200 OK
+(EngineCore pid=312907) INFO 07-15 19:47:50 [core.py:1214] [shutdown] EngineCore: trigger received signal=SIGTERM
+(APIServer pid=312789) INFO 07-15 19:47:50 [launcher.py:100] [shutdown] API server: shutdown triggered
+(APIServer pid=312789) INFO 07-15 19:47:50 [launcher.py:116] [shutdown] API server: stopping engine client mode=abort timeout=0s
+(EngineCore pid=312907) INFO 07-15 19:47:50 [core.py:1333] [shutdown] EngineCore: start mode=abort timeout=0s
+(EngineCore pid=312907) INFO 07-15 19:47:50 [core.py:1364] [shutdown] EngineCore: request processing complete; starting resource teardown
+(EngineCore pid=312907) INFO 07-15 19:47:50 [core.py:1227] [shutdown] EngineCore: exiting busy loop
+(APIServer pid=312789) INFO 07-15 19:47:50 [core_client.py:655] [shutdown] MPClient: start timeout=0s
+(APIServer pid=312789) INFO 07-15 19:47:50 [core_client.py:657] [shutdown] MPClient: stopping engine manager
+(APIServer pid=312789) WARNING 07-15 19:47:50 [utils.py:626] [shutdown] Process manager: force killing remaining processes count=1
+(APIServer pid=312789) INFO 07-15 19:47:50 [core_client.py:659] [shutdown] MPClient: engine manager stopped
+(APIServer pid=312789) INFO 07-15 19:47:50 [core_client.py:660] [shutdown] MPClient: cleaning up background resources
+(APIServer pid=312789) INFO 07-15 19:47:50 [core_client.py:662] [shutdown] MPClient: complete
+(APIServer pid=312789) INFO 07-15 19:47:50 [launcher.py:125] [shutdown] API server: engine client stopped
+(APIServer pid=312789) INFO 07-15 19:47:50 [launcher.py:128] [shutdown] API server: signalling HTTP server shutdown
+(APIServer pid=312789) INFO 07-15 19:47:50 [launcher.py:149] [shutdown] API server: shutting down FastAPI HTTP server
+(APIServer pid=312789) INFO:     Shutting down
+(APIServer pid=312789) INFO:     Waiting for application shutdown.
+(APIServer pid=312789) INFO:     Application shutdown complete.
+/home/henry/.local/share/uv/python/cpython-3.12.12-linux-x86_64-gnu/lib/python3.12/multiprocessing/resource_tracker.py:279: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
+  warnings.warn('resource_tracker: There appear to be %d '

evals/healing_breadth/keep25_traj/step0100_chat.json.server.log

@@ -0,0 +1,134 @@
+(APIServer pid=313816) INFO 07-15 19:48:14 [api_utils.py:339] 
+(APIServer pid=313816) INFO 07-15 19:48:14 [api_utils.py:339]        █     █     █▄   ▄█
+(APIServer pid=313816) INFO 07-15 19:48:14 [api_utils.py:339]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.25.0
+(APIServer pid=313816) INFO 07-15 19:48:14 [api_utils.py:339]   █▄█▀ █     █     █     █  model   outputs/healed/healing_breadth/glean_math_keep25_seed1224_long/step0100
+(APIServer pid=313816) INFO 07-15 19:48:14 [api_utils.py:339]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
+(APIServer pid=313816) INFO 07-15 19:48:14 [api_utils.py:339] 
+(APIServer pid=313816) INFO 07-15 19:48:14 [api_utils.py:273] non-default args: {'model_tag': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long/step0100', 'host': '127.0.0.1', 'port': 8377, 'model': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long/step0100', 'max_model_len': 2048, 'enforce_eager': True, 'served_model_name': ['student'], 'gpu_memory_utilization': 0.85}
+(APIServer pid=313816) INFO 07-15 19:48:14 [model.py:619] Resolved architecture: PrunedOlmoeForCausalLM
+(APIServer pid=313816) INFO 07-15 19:48:14 [model.py:1776] Using max model len 2048
+(APIServer pid=313816) INFO 07-15 19:48:14 [vllm.py:1042] Asynchronous scheduling is enabled.
+(APIServer pid=313816) WARNING 07-15 19:48:14 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(APIServer pid=313816) WARNING 07-15 19:48:14 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(APIServer pid=313816) INFO 07-15 19:48:14 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(APIServer pid=313816) INFO 07-15 19:48:14 [vllm.py:1322] Cudagraph is disabled under eager mode
+(APIServer pid=313816) INFO 07-15 19:48:14 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(EngineCore pid=313953) INFO 07-15 19:48:21 [core.py:114] Initializing a V1 LLM engine (v0.25.0) with config: model='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long/step0100', speculative_config=None, tokenizer='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long/step0100', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=student, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.NONE: 0>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, moe_backend='auto', linear_backend='auto')
+(EngineCore pid=313953) INFO 07-15 19:48:22 [parallel_state.py:1607] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.0.15:35799 backend=nccl
+(EngineCore pid=313953) INFO 07-15 19:48:22 [parallel_state.py:1942] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A
+(EngineCore pid=313953) INFO 07-15 19:48:23 [topk_topp_sampler.py:55] Using FlashInfer for top-p & top-k sampling.
+(EngineCore pid=313953) INFO 07-15 19:48:23 [gpu_model_runner.py:5209] Starting to load model outputs/healed/healing_breadth/glean_math_keep25_seed1224_long/step0100...
+(EngineCore pid=313953) INFO 07-15 19:48:23 [cuda.py:476] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
+(EngineCore pid=313953) INFO 07-15 19:48:23 [flash_attn.py:718] Using FlashAttention version 2
+(EngineCore pid=313953) /home/henry/.cache/glean/megablocks-variable-93a1479bc15b/megablocks/grouped_gemm_util.py:10: UserWarning: Grouped GEMM not available.
+(EngineCore pid=313953)   warnings.warn('Grouped GEMM not available.')
+(EngineCore pid=313953) INFO 07-15 19:48:24 [weight_utils.py:849] Filesystem type for checkpoints: EXT4. Checkpoint size: 3.89 GiB. Available RAM: 109.26 GiB.
+(EngineCore pid=313953) INFO 07-15 19:48:24 [weight_utils.py:872] Auto-prefetch is disabled because the filesystem (EXT4) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.
+(EngineCore pid=313953) 
+(EngineCore pid=313953) 
+(EngineCore pid=313953) 
+(EngineCore pid=313953) 
+(EngineCore pid=313953) INFO 07-15 19:48:30 [default_loader.py:430] Loading weights took 6.22 seconds
+(EngineCore pid=313953) INFO 07-15 19:48:30 [gpu_model_runner.py:5306] Model loading took 3.89 GiB memory and 6.405053 seconds
+(EngineCore pid=313953) INFO 07-15 19:48:32 [gpu_worker.py:538] Available KV cache memory: 15.82 GiB
+(EngineCore pid=313953) INFO 07-15 19:48:32 [kv_cache_utils.py:2146] GPU KV cache size: 129,584 tokens
+(EngineCore pid=313953) INFO 07-15 19:48:32 [kv_cache_utils.py:2147] Maximum concurrency for 2,048 tokens per request: 63.27x
+(EngineCore pid=313953) INFO 07-15 19:48:32 [cutedsl_warmup.py:97] Skipping CuTeDSL warmup because no compile units were requested.
+(EngineCore pid=313953) INFO 07-15 19:48:32 [jit_monitor.py:73] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
+(EngineCore pid=313953) INFO 07-15 19:48:32 [core.py:344] init engine (profile, create kv cache, warmup model) took 2.06 s
+(EngineCore pid=313953) INFO 07-15 19:48:32 [vllm.py:1042] Asynchronous scheduling is enabled.
+(EngineCore pid=313953) WARNING 07-15 19:48:32 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(EngineCore pid=313953) WARNING 07-15 19:48:32 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(EngineCore pid=313953) INFO 07-15 19:48:32 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(EngineCore pid=313953) INFO 07-15 19:48:32 [vllm.py:1322] Cudagraph is disabled under eager mode
+(EngineCore pid=313953) INFO 07-15 19:48:32 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(APIServer pid=313816) INFO 07-15 19:48:32 [api_server.py:612] Supported tasks: ['generate']
+(APIServer pid=313816) WARNING 07-15 19:48:33 [__init__.py:36] SECURITY WARNING: Development endpoints are enabled! This should NOT be used in production!
+(APIServer pid=313816) INFO 07-15 19:48:33 [hf.py:548] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
+(APIServer pid=313816) INFO 07-15 19:48:33 [api_server.py:616] Starting vLLM server on http://127.0.0.1:8377
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:37] Available routes are:
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /openapi.json, Methods: HEAD, GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /docs, Methods: HEAD, GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: HEAD, GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /redoc, Methods: HEAD, GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /load, Methods: GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /version, Methods: GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /health, Methods: GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /metrics, Methods: GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /tokenize, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /detokenize, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /v1/models, Methods: GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /ping, Methods: GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /ping, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /invocations, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /reset_prefix_cache, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /reset_mm_cache, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /reset_encoder_cache, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /pause, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /resume, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /is_paused, Methods: GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /init_weight_transfer_engine, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /start_weight_update, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /update_weights, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /finish_weight_update, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /get_world_size, Methods: GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /collective_rpc, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /server_info, Methods: GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /sleep, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /wake_up, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /is_sleeping, Methods: GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /v1/chat/completions/batch, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /v1/responses, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /v1/completions, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /v1/messages, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /generative_scoring, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /v1/completions/render, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /v1/chat/completions/derender, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /v1/completions/derender, Methods: POST
+(APIServer pid=313816) INFO 07-15 19:48:33 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
+(APIServer pid=313816) INFO:     Started server process [313816]
+(APIServer pid=313816) INFO:     Waiting for application startup.
+(APIServer pid=313816) INFO:     Application startup complete.
+(APIServer pid=313816) INFO:     127.0.0.1:33030 - "GET /health HTTP/1.1" 200 OK
+(EngineCore pid=313953) WARNING 07-15 19:48:35 [jit_monitor.py:129] Triton kernel JIT compilation during inference: _build_route_rows. This causes a latency spike; consider extending warmup to cover this shape/config.
+(APIServer pid=313816) INFO 07-15 19:48:43 [loggers.py:273] Engine 000: Avg prompt throughput: 2065.1 tokens/s, Avg generation throughput: 2145.0 tokens/s, Running: 256 reqs, Waiting: 1056 reqs, GPU KV cache usage: 35.1%, Prefix cache hit rate: 93.1%
+(APIServer pid=313816) INFO 07-15 19:48:53 [loggers.py:273] Engine 000: Avg prompt throughput: 563.3 tokens/s, Avg generation throughput: 2962.1 tokens/s, Running: 256 reqs, Waiting: 983 reqs, GPU KV cache usage: 49.2%, Prefix cache hit rate: 93.2%
+(APIServer pid=313816) INFO 07-15 19:49:03 [loggers.py:273] Engine 000: Avg prompt throughput: 861.6 tokens/s, Avg generation throughput: 2906.9 tokens/s, Running: 255 reqs, Waiting: 875 reqs, GPU KV cache usage: 50.5%, Prefix cache hit rate: 93.2%
+(APIServer pid=313816) INFO 07-15 19:49:13 [loggers.py:273] Engine 000: Avg prompt throughput: 764.1 tokens/s, Avg generation throughput: 2882.7 tokens/s, Running: 255 reqs, Waiting: 775 reqs, GPU KV cache usage: 52.4%, Prefix cache hit rate: 93.3%
+(APIServer pid=313816) INFO 07-15 19:49:23 [loggers.py:273] Engine 000: Avg prompt throughput: 1012.0 tokens/s, Avg generation throughput: 2905.1 tokens/s, Running: 255 reqs, Waiting: 644 reqs, GPU KV cache usage: 43.6%, Prefix cache hit rate: 93.3%
+(APIServer pid=313816) INFO 07-15 19:49:33 [loggers.py:273] Engine 000: Avg prompt throughput: 738.2 tokens/s, Avg generation throughput: 2934.5 tokens/s, Running: 256 reqs, Waiting: 551 reqs, GPU KV cache usage: 48.9%, Prefix cache hit rate: 93.4%
+(APIServer pid=313816) INFO 07-15 19:49:43 [loggers.py:273] Engine 000: Avg prompt throughput: 907.9 tokens/s, Avg generation throughput: 2906.6 tokens/s, Running: 255 reqs, Waiting: 437 reqs, GPU KV cache usage: 47.7%, Prefix cache hit rate: 93.3%
+(APIServer pid=313816) INFO 07-15 19:49:53 [loggers.py:273] Engine 000: Avg prompt throughput: 820.2 tokens/s, Avg generation throughput: 2934.0 tokens/s, Running: 254 reqs, Waiting: 336 reqs, GPU KV cache usage: 50.0%, Prefix cache hit rate: 93.3%
+(APIServer pid=313816) INFO 07-15 19:50:03 [loggers.py:273] Engine 000: Avg prompt throughput: 953.8 tokens/s, Avg generation throughput: 2906.3 tokens/s, Running: 255 reqs, Waiting: 219 reqs, GPU KV cache usage: 47.8%, Prefix cache hit rate: 93.4%
+(APIServer pid=313816) INFO 07-15 19:50:13 [loggers.py:273] Engine 000: Avg prompt throughput: 705.9 tokens/s, Avg generation throughput: 2934.8 tokens/s, Running: 253 reqs, Waiting: 126 reqs, GPU KV cache usage: 50.6%, Prefix cache hit rate: 93.4%
+(APIServer pid=313816) INFO 07-15 19:50:23 [loggers.py:273] Engine 000: Avg prompt throughput: 946.6 tokens/s, Avg generation throughput: 2881.4 tokens/s, Running: 255 reqs, Waiting: 14 reqs, GPU KV cache usage: 49.7%, Prefix cache hit rate: 93.4%
+(APIServer pid=313816) INFO 07-15 19:50:33 [loggers.py:273] Engine 000: Avg prompt throughput: 108.3 tokens/s, Avg generation throughput: 2797.2 tokens/s, Running: 158 reqs, Waiting: 0 reqs, GPU KV cache usage: 40.4%, Prefix cache hit rate: 93.4%
+(APIServer pid=313816) INFO 07-15 19:50:43 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1614.5 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 93.4%
+(APIServer pid=313816) INFO:     127.0.0.1:33038 - "POST /v1/completions HTTP/1.1" 200 OK
+(EngineCore pid=313953) INFO 07-15 19:50:46 [core.py:1214] [shutdown] EngineCore: trigger received signal=SIGTERM
+(APIServer pid=313816) INFO 07-15 19:50:46 [launcher.py:100] [shutdown] API server: shutdown triggered
+(APIServer pid=313816) INFO 07-15 19:50:46 [launcher.py:116] [shutdown] API server: stopping engine client mode=abort timeout=0s
+(EngineCore pid=313953) INFO 07-15 19:50:46 [core.py:1333] [shutdown] EngineCore: start mode=abort timeout=0s
+(EngineCore pid=313953) INFO 07-15 19:50:46 [core.py:1364] [shutdown] EngineCore: request processing complete; starting resource teardown
+(EngineCore pid=313953) INFO 07-15 19:50:46 [core.py:1227] [shutdown] EngineCore: exiting busy loop
+(APIServer pid=313816) INFO 07-15 19:50:46 [core_client.py:655] [shutdown] MPClient: start timeout=0s
+(APIServer pid=313816) INFO 07-15 19:50:46 [core_client.py:657] [shutdown] MPClient: stopping engine manager
+(APIServer pid=313816) WARNING 07-15 19:50:46 [utils.py:626] [shutdown] Process manager: force killing remaining processes count=1
+(APIServer pid=313816) INFO 07-15 19:50:46 [core_client.py:659] [shutdown] MPClient: engine manager stopped
+(APIServer pid=313816) INFO 07-15 19:50:46 [core_client.py:660] [shutdown] MPClient: cleaning up background resources
+(APIServer pid=313816) INFO 07-15 19:50:46 [core_client.py:662] [shutdown] MPClient: complete
+(APIServer pid=313816) INFO 07-15 19:50:46 [launcher.py:125] [shutdown] API server: engine client stopped
+(APIServer pid=313816) INFO 07-15 19:50:46 [launcher.py:128] [shutdown] API server: signalling HTTP server shutdown
+(APIServer pid=313816) INFO 07-15 19:50:46 [launcher.py:149] [shutdown] API server: shutting down FastAPI HTTP server
+(APIServer pid=313816) INFO:     Shutting down
+(APIServer pid=313816) INFO:     Waiting for application shutdown.
+(APIServer pid=313816) INFO:     Application shutdown complete.
+/home/henry/.local/share/uv/python/cpython-3.12.12-linux-x86_64-gnu/lib/python3.12/multiprocessing/resource_tracker.py:279: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
+  warnings.warn('resource_tracker: There appear to be %d '

evals/healing_breadth/keep25_traj/step0150_chat.json.server.log

@@ -0,0 +1,133 @@
+(APIServer pid=314661) INFO 07-15 19:51:11 [api_utils.py:339] 
+(APIServer pid=314661) INFO 07-15 19:51:11 [api_utils.py:339]        █     █     █▄   ▄█
+(APIServer pid=314661) INFO 07-15 19:51:11 [api_utils.py:339]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.25.0
+(APIServer pid=314661) INFO 07-15 19:51:11 [api_utils.py:339]   █▄█▀ █     █     █     █  model   outputs/healed/healing_breadth/glean_math_keep25_seed1224_long/step0150
+(APIServer pid=314661) INFO 07-15 19:51:11 [api_utils.py:339]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
+(APIServer pid=314661) INFO 07-15 19:51:11 [api_utils.py:339] 
+(APIServer pid=314661) INFO 07-15 19:51:11 [api_utils.py:273] non-default args: {'model_tag': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long/step0150', 'host': '127.0.0.1', 'port': 8377, 'model': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long/step0150', 'max_model_len': 2048, 'enforce_eager': True, 'served_model_name': ['student'], 'gpu_memory_utilization': 0.85}
+(APIServer pid=314661) INFO 07-15 19:51:11 [model.py:619] Resolved architecture: PrunedOlmoeForCausalLM
+(APIServer pid=314661) INFO 07-15 19:51:11 [model.py:1776] Using max model len 2048
+(APIServer pid=314661) INFO 07-15 19:51:11 [vllm.py:1042] Asynchronous scheduling is enabled.
+(APIServer pid=314661) WARNING 07-15 19:51:11 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(APIServer pid=314661) WARNING 07-15 19:51:11 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(APIServer pid=314661) INFO 07-15 19:51:11 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(APIServer pid=314661) INFO 07-15 19:51:11 [vllm.py:1322] Cudagraph is disabled under eager mode
+(APIServer pid=314661) INFO 07-15 19:51:11 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(EngineCore pid=314779) INFO 07-15 19:51:18 [core.py:114] Initializing a V1 LLM engine (v0.25.0) with config: model='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long/step0150', speculative_config=None, tokenizer='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long/step0150', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=student, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.NONE: 0>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, moe_backend='auto', linear_backend='auto')
+(EngineCore pid=314779) INFO 07-15 19:51:19 [parallel_state.py:1607] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.0.15:48423 backend=nccl
+(EngineCore pid=314779) INFO 07-15 19:51:19 [parallel_state.py:1942] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A
+(EngineCore pid=314779) INFO 07-15 19:51:19 [topk_topp_sampler.py:55] Using FlashInfer for top-p & top-k sampling.
+(EngineCore pid=314779) INFO 07-15 19:51:20 [gpu_model_runner.py:5209] Starting to load model outputs/healed/healing_breadth/glean_math_keep25_seed1224_long/step0150...
+(EngineCore pid=314779) INFO 07-15 19:51:20 [cuda.py:476] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
+(EngineCore pid=314779) INFO 07-15 19:51:20 [flash_attn.py:718] Using FlashAttention version 2
+(EngineCore pid=314779) /home/henry/.cache/glean/megablocks-variable-93a1479bc15b/megablocks/grouped_gemm_util.py:10: UserWarning: Grouped GEMM not available.
+(EngineCore pid=314779)   warnings.warn('Grouped GEMM not available.')
+(EngineCore pid=314779) INFO 07-15 19:51:20 [weight_utils.py:849] Filesystem type for checkpoints: EXT4. Checkpoint size: 3.89 GiB. Available RAM: 109.07 GiB.
+(EngineCore pid=314779) INFO 07-15 19:51:20 [weight_utils.py:872] Auto-prefetch is disabled because the filesystem (EXT4) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.
+(EngineCore pid=314779) 
+(EngineCore pid=314779) 
+(EngineCore pid=314779) 
+(EngineCore pid=314779) 
+(EngineCore pid=314779) INFO 07-15 19:51:26 [default_loader.py:430] Loading weights took 6.46 seconds
+(EngineCore pid=314779) INFO 07-15 19:51:27 [gpu_model_runner.py:5306] Model loading took 3.89 GiB memory and 6.638648 seconds
+(EngineCore pid=314779) INFO 07-15 19:51:28 [gpu_worker.py:538] Available KV cache memory: 15.82 GiB
+(EngineCore pid=314779) INFO 07-15 19:51:28 [kv_cache_utils.py:2146] GPU KV cache size: 129,584 tokens
+(EngineCore pid=314779) INFO 07-15 19:51:28 [kv_cache_utils.py:2147] Maximum concurrency for 2,048 tokens per request: 63.27x
+(EngineCore pid=314779) INFO 07-15 19:51:29 [cutedsl_warmup.py:97] Skipping CuTeDSL warmup because no compile units were requested.
+(EngineCore pid=314779) INFO 07-15 19:51:29 [jit_monitor.py:73] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
+(EngineCore pid=314779) INFO 07-15 19:51:29 [core.py:344] init engine (profile, create kv cache, warmup model) took 2.06 s
+(EngineCore pid=314779) INFO 07-15 19:51:29 [vllm.py:1042] Asynchronous scheduling is enabled.
+(EngineCore pid=314779) WARNING 07-15 19:51:29 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(EngineCore pid=314779) WARNING 07-15 19:51:29 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(EngineCore pid=314779) INFO 07-15 19:51:29 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(EngineCore pid=314779) INFO 07-15 19:51:29 [vllm.py:1322] Cudagraph is disabled under eager mode
+(EngineCore pid=314779) INFO 07-15 19:51:29 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(APIServer pid=314661) INFO 07-15 19:51:29 [api_server.py:612] Supported tasks: ['generate']
+(APIServer pid=314661) WARNING 07-15 19:51:29 [__init__.py:36] SECURITY WARNING: Development endpoints are enabled! This should NOT be used in production!
+(APIServer pid=314661) INFO 07-15 19:51:29 [hf.py:548] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
+(APIServer pid=314661) INFO 07-15 19:51:29 [api_server.py:616] Starting vLLM server on http://127.0.0.1:8377
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:37] Available routes are:
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /openapi.json, Methods: GET, HEAD
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /docs, Methods: GET, HEAD
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: GET, HEAD
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /redoc, Methods: GET, HEAD
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /load, Methods: GET
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /version, Methods: GET
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /health, Methods: GET
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /metrics, Methods: GET
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /tokenize, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /detokenize, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /v1/models, Methods: GET
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /ping, Methods: GET
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /ping, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /invocations, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /reset_prefix_cache, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /reset_mm_cache, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /reset_encoder_cache, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /pause, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /resume, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /is_paused, Methods: GET
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /init_weight_transfer_engine, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /start_weight_update, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /update_weights, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /finish_weight_update, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /get_world_size, Methods: GET
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /collective_rpc, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /server_info, Methods: GET
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /sleep, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /wake_up, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /is_sleeping, Methods: GET
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /v1/chat/completions/batch, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /v1/responses, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /v1/completions, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /v1/messages, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /generative_scoring, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /v1/completions/render, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /v1/chat/completions/derender, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /v1/completions/derender, Methods: POST
+(APIServer pid=314661) INFO 07-15 19:51:29 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
+(APIServer pid=314661) INFO:     Started server process [314661]
+(APIServer pid=314661) INFO:     Waiting for application startup.
+(APIServer pid=314661) INFO:     Application startup complete.
+(APIServer pid=314661) INFO:     127.0.0.1:46516 - "GET /health HTTP/1.1" 200 OK
+(EngineCore pid=314779) WARNING 07-15 19:51:31 [jit_monitor.py:129] Triton kernel JIT compilation during inference: _build_route_rows. This causes a latency spike; consider extending warmup to cover this shape/config.
+(APIServer pid=314661) INFO 07-15 19:51:40 [loggers.py:273] Engine 000: Avg prompt throughput: 2054.6 tokens/s, Avg generation throughput: 2224.5 tokens/s, Running: 256 reqs, Waiting: 1058 reqs, GPU KV cache usage: 35.8%, Prefix cache hit rate: 93.1%
+(APIServer pid=314661) INFO 07-15 19:51:50 [loggers.py:273] Engine 000: Avg prompt throughput: 645.2 tokens/s, Avg generation throughput: 2960.7 tokens/s, Running: 255 reqs, Waiting: 973 reqs, GPU KV cache usage: 47.7%, Prefix cache hit rate: 93.2%
+(APIServer pid=314661) INFO 07-15 19:52:00 [loggers.py:273] Engine 000: Avg prompt throughput: 1064.6 tokens/s, Avg generation throughput: 2904.9 tokens/s, Running: 255 reqs, Waiting: 840 reqs, GPU KV cache usage: 45.4%, Prefix cache hit rate: 93.2%
+(APIServer pid=314661) INFO 07-15 19:52:10 [loggers.py:273] Engine 000: Avg prompt throughput: 723.5 tokens/s, Avg generation throughput: 2933.9 tokens/s, Running: 254 reqs, Waiting: 743 reqs, GPU KV cache usage: 49.8%, Prefix cache hit rate: 93.3%
+(APIServer pid=314661) INFO 07-15 19:52:20 [loggers.py:273] Engine 000: Avg prompt throughput: 950.4 tokens/s, Avg generation throughput: 2906.3 tokens/s, Running: 256 reqs, Waiting: 623 reqs, GPU KV cache usage: 46.6%, Prefix cache hit rate: 93.3%
+(APIServer pid=314661) INFO 07-15 19:52:30 [loggers.py:273] Engine 000: Avg prompt throughput: 980.8 tokens/s, Avg generation throughput: 2931.5 tokens/s, Running: 254 reqs, Waiting: 501 reqs, GPU KV cache usage: 45.5%, Prefix cache hit rate: 93.3%
+(APIServer pid=314661) INFO 07-15 19:52:40 [loggers.py:273] Engine 000: Avg prompt throughput: 832.0 tokens/s, Avg generation throughput: 2933.1 tokens/s, Running: 255 reqs, Waiting: 393 reqs, GPU KV cache usage: 47.1%, Prefix cache hit rate: 93.4%
+(APIServer pid=314661) INFO 07-15 19:52:50 [loggers.py:273] Engine 000: Avg prompt throughput: 1027.4 tokens/s, Avg generation throughput: 2905.8 tokens/s, Running: 255 reqs, Waiting: 269 reqs, GPU KV cache usage: 46.3%, Prefix cache hit rate: 93.4%
+(APIServer pid=314661) INFO 07-15 19:53:00 [loggers.py:273] Engine 000: Avg prompt throughput: 924.7 tokens/s, Avg generation throughput: 2932.1 tokens/s, Running: 254 reqs, Waiting: 150 reqs, GPU KV cache usage: 45.7%, Prefix cache hit rate: 93.4%
+(APIServer pid=314661) INFO 07-15 19:53:10 [loggers.py:273] Engine 000: Avg prompt throughput: 777.6 tokens/s, Avg generation throughput: 2934.4 tokens/s, Running: 256 reqs, Waiting: 58 reqs, GPU KV cache usage: 52.3%, Prefix cache hit rate: 93.4%
+(APIServer pid=314661) INFO 07-15 19:53:20 [loggers.py:273] Engine 000: Avg prompt throughput: 466.7 tokens/s, Avg generation throughput: 2847.6 tokens/s, Running: 184 reqs, Waiting: 0 reqs, GPU KV cache usage: 40.9%, Prefix cache hit rate: 93.4%
+(APIServer pid=314661) INFO 07-15 19:53:30 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2074.7 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.8%, Prefix cache hit rate: 93.4%
+(APIServer pid=314661) INFO:     127.0.0.1:46530 - "POST /v1/completions HTTP/1.1" 200 OK
+(EngineCore pid=314779) INFO 07-15 19:53:36 [core.py:1214] [shutdown] EngineCore: trigger received signal=SIGTERM
+(APIServer pid=314661) INFO 07-15 19:53:36 [launcher.py:100] [shutdown] API server: shutdown triggered
+(APIServer pid=314661) INFO 07-15 19:53:36 [launcher.py:116] [shutdown] API server: stopping engine client mode=abort timeout=0s
+(EngineCore pid=314779) INFO 07-15 19:53:36 [core.py:1333] [shutdown] EngineCore: start mode=abort timeout=0s
+(EngineCore pid=314779) INFO 07-15 19:53:36 [core.py:1364] [shutdown] EngineCore: request processing complete; starting resource teardown
+(EngineCore pid=314779) INFO 07-15 19:53:36 [core.py:1227] [shutdown] EngineCore: exiting busy loop
+(APIServer pid=314661) INFO 07-15 19:53:36 [core_client.py:655] [shutdown] MPClient: start timeout=0s
+(APIServer pid=314661) INFO 07-15 19:53:36 [core_client.py:657] [shutdown] MPClient: stopping engine manager
+(APIServer pid=314661) WARNING 07-15 19:53:36 [utils.py:626] [shutdown] Process manager: force killing remaining processes count=1
+(APIServer pid=314661) INFO 07-15 19:53:36 [core_client.py:659] [shutdown] MPClient: engine manager stopped
+(APIServer pid=314661) INFO 07-15 19:53:36 [core_client.py:660] [shutdown] MPClient: cleaning up background resources
+(APIServer pid=314661) INFO 07-15 19:53:36 [core_client.py:662] [shutdown] MPClient: complete
+(APIServer pid=314661) INFO 07-15 19:53:36 [launcher.py:125] [shutdown] API server: engine client stopped
+(APIServer pid=314661) INFO 07-15 19:53:36 [launcher.py:128] [shutdown] API server: signalling HTTP server shutdown
+(APIServer pid=314661) INFO 07-15 19:53:36 [launcher.py:149] [shutdown] API server: shutting down FastAPI HTTP server
+(APIServer pid=314661) INFO:     Shutting down
+(APIServer pid=314661) INFO:     Waiting for application shutdown.
+(APIServer pid=314661) INFO:     Application shutdown complete.
+/home/henry/.local/share/uv/python/cpython-3.12.12-linux-x86_64-gnu/lib/python3.12/multiprocessing/resource_tracker.py:279: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
+  warnings.warn('resource_tracker: There appear to be %d '

evals/healing_breadth/keep25_traj/step0200_chat.json.server.log

@@ -0,0 +1,133 @@
+(APIServer pid=316100) INFO 07-15 19:54:01 [api_utils.py:339] 
+(APIServer pid=316100) INFO 07-15 19:54:01 [api_utils.py:339]        █     █     █▄   ▄█
+(APIServer pid=316100) INFO 07-15 19:54:01 [api_utils.py:339]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.25.0
+(APIServer pid=316100) INFO 07-15 19:54:01 [api_utils.py:339]   █▄█▀ █     █     █     █  model   outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0200
+(APIServer pid=316100) INFO 07-15 19:54:01 [api_utils.py:339]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
+(APIServer pid=316100) INFO 07-15 19:54:01 [api_utils.py:339] 
+(APIServer pid=316100) INFO 07-15 19:54:01 [api_utils.py:273] non-default args: {'model_tag': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0200', 'host': '127.0.0.1', 'port': 8377, 'model': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0200', 'max_model_len': 2048, 'enforce_eager': True, 'served_model_name': ['student'], 'gpu_memory_utilization': 0.85}
+(APIServer pid=316100) INFO 07-15 19:54:01 [model.py:619] Resolved architecture: PrunedOlmoeForCausalLM
+(APIServer pid=316100) INFO 07-15 19:54:01 [model.py:1776] Using max model len 2048
+(APIServer pid=316100) INFO 07-15 19:54:01 [vllm.py:1042] Asynchronous scheduling is enabled.
+(APIServer pid=316100) WARNING 07-15 19:54:01 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(APIServer pid=316100) WARNING 07-15 19:54:01 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(APIServer pid=316100) INFO 07-15 19:54:01 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(APIServer pid=316100) INFO 07-15 19:54:01 [vllm.py:1322] Cudagraph is disabled under eager mode
+(APIServer pid=316100) INFO 07-15 19:54:01 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(EngineCore pid=316217) INFO 07-15 19:54:08 [core.py:114] Initializing a V1 LLM engine (v0.25.0) with config: model='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0200', speculative_config=None, tokenizer='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0200', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=student, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.NONE: 0>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, moe_backend='auto', linear_backend='auto')
+(EngineCore pid=316217) INFO 07-15 19:54:09 [parallel_state.py:1607] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.0.15:45881 backend=nccl
+(EngineCore pid=316217) INFO 07-15 19:54:09 [parallel_state.py:1942] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A
+(EngineCore pid=316217) INFO 07-15 19:54:10 [topk_topp_sampler.py:55] Using FlashInfer for top-p & top-k sampling.
+(EngineCore pid=316217) INFO 07-15 19:54:10 [gpu_model_runner.py:5209] Starting to load model outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0200...
+(EngineCore pid=316217) INFO 07-15 19:54:10 [cuda.py:476] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
+(EngineCore pid=316217) INFO 07-15 19:54:10 [flash_attn.py:718] Using FlashAttention version 2
+(EngineCore pid=316217) /home/henry/.cache/glean/megablocks-variable-93a1479bc15b/megablocks/grouped_gemm_util.py:10: UserWarning: Grouped GEMM not available.
+(EngineCore pid=316217)   warnings.warn('Grouped GEMM not available.')
+(EngineCore pid=316217) INFO 07-15 19:54:10 [weight_utils.py:849] Filesystem type for checkpoints: EXT4. Checkpoint size: 3.89 GiB. Available RAM: 106.70 GiB.
+(EngineCore pid=316217) INFO 07-15 19:54:10 [weight_utils.py:872] Auto-prefetch is disabled because the filesystem (EXT4) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.
+(EngineCore pid=316217) 
+(EngineCore pid=316217) 
+(EngineCore pid=316217) 
+(EngineCore pid=316217) 
+(EngineCore pid=316217) INFO 07-15 19:54:12 [default_loader.py:430] Loading weights took 2.41 seconds
+(EngineCore pid=316217) INFO 07-15 19:54:13 [gpu_model_runner.py:5306] Model loading took 3.89 GiB memory and 2.596651 seconds
+(EngineCore pid=316217) INFO 07-15 19:54:15 [gpu_worker.py:538] Available KV cache memory: 15.82 GiB
+(EngineCore pid=316217) INFO 07-15 19:54:15 [kv_cache_utils.py:2146] GPU KV cache size: 129,584 tokens
+(EngineCore pid=316217) INFO 07-15 19:54:15 [kv_cache_utils.py:2147] Maximum concurrency for 2,048 tokens per request: 63.27x
+(EngineCore pid=316217) INFO 07-15 19:54:15 [cutedsl_warmup.py:97] Skipping CuTeDSL warmup because no compile units were requested.
+(EngineCore pid=316217) INFO 07-15 19:54:15 [jit_monitor.py:73] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
+(EngineCore pid=316217) INFO 07-15 19:54:15 [core.py:344] init engine (profile, create kv cache, warmup model) took 2.11 s
+(EngineCore pid=316217) INFO 07-15 19:54:15 [vllm.py:1042] Asynchronous scheduling is enabled.
+(EngineCore pid=316217) WARNING 07-15 19:54:15 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(EngineCore pid=316217) WARNING 07-15 19:54:15 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(EngineCore pid=316217) INFO 07-15 19:54:15 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(EngineCore pid=316217) INFO 07-15 19:54:15 [vllm.py:1322] Cudagraph is disabled under eager mode
+(EngineCore pid=316217) INFO 07-15 19:54:15 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(APIServer pid=316100) INFO 07-15 19:54:15 [api_server.py:612] Supported tasks: ['generate']
+(APIServer pid=316100) WARNING 07-15 19:54:15 [__init__.py:36] SECURITY WARNING: Development endpoints are enabled! This should NOT be used in production!
+(APIServer pid=316100) INFO 07-15 19:54:16 [hf.py:548] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
+(APIServer pid=316100) INFO 07-15 19:54:16 [api_server.py:616] Starting vLLM server on http://127.0.0.1:8377
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:37] Available routes are:
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /openapi.json, Methods: HEAD, GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /docs, Methods: HEAD, GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: HEAD, GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /redoc, Methods: HEAD, GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /load, Methods: GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /version, Methods: GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /health, Methods: GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /metrics, Methods: GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /tokenize, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /detokenize, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /v1/models, Methods: GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /ping, Methods: GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /ping, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /invocations, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /reset_prefix_cache, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /reset_mm_cache, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /reset_encoder_cache, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /pause, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /resume, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /is_paused, Methods: GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /init_weight_transfer_engine, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /start_weight_update, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /update_weights, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /finish_weight_update, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /get_world_size, Methods: GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /collective_rpc, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /server_info, Methods: GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /sleep, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /wake_up, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /is_sleeping, Methods: GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /v1/chat/completions/batch, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /v1/responses, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /v1/completions, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /v1/messages, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /generative_scoring, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /v1/completions/render, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /v1/chat/completions/derender, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /v1/completions/derender, Methods: POST
+(APIServer pid=316100) INFO 07-15 19:54:16 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
+(APIServer pid=316100) INFO:     Started server process [316100]
+(APIServer pid=316100) INFO:     Waiting for application startup.
+(APIServer pid=316100) INFO:     Application startup complete.
+(APIServer pid=316100) INFO:     127.0.0.1:38268 - "GET /health HTTP/1.1" 200 OK
+(EngineCore pid=316217) WARNING 07-15 19:54:17 [jit_monitor.py:129] Triton kernel JIT compilation during inference: _build_route_rows. This causes a latency spike; consider extending warmup to cover this shape/config.
+(APIServer pid=316100) INFO 07-15 19:54:26 [loggers.py:273] Engine 000: Avg prompt throughput: 2080.7 tokens/s, Avg generation throughput: 2245.3 tokens/s, Running: 256 reqs, Waiting: 1053 reqs, GPU KV cache usage: 35.7%, Prefix cache hit rate: 93.1%
+(APIServer pid=316100) INFO 07-15 19:54:36 [loggers.py:273] Engine 000: Avg prompt throughput: 800.1 tokens/s, Avg generation throughput: 2958.9 tokens/s, Running: 255 reqs, Waiting: 953 reqs, GPU KV cache usage: 46.0%, Prefix cache hit rate: 93.2%
+(APIServer pid=316100) INFO 07-15 19:54:46 [loggers.py:273] Engine 000: Avg prompt throughput: 924.1 tokens/s, Avg generation throughput: 2906.6 tokens/s, Running: 256 reqs, Waiting: 836 reqs, GPU KV cache usage: 47.4%, Prefix cache hit rate: 93.2%
+(APIServer pid=316100) INFO 07-15 19:54:56 [loggers.py:273] Engine 000: Avg prompt throughput: 826.9 tokens/s, Avg generation throughput: 2907.3 tokens/s, Running: 253 reqs, Waiting: 726 reqs, GPU KV cache usage: 47.9%, Prefix cache hit rate: 93.3%
+(APIServer pid=316100) INFO 07-15 19:55:06 [loggers.py:273] Engine 000: Avg prompt throughput: 922.4 tokens/s, Avg generation throughput: 2906.3 tokens/s, Running: 256 reqs, Waiting: 607 reqs, GPU KV cache usage: 45.5%, Prefix cache hit rate: 93.3%
+(APIServer pid=316100) INFO 07-15 19:55:16 [loggers.py:273] Engine 000: Avg prompt throughput: 1067.3 tokens/s, Avg generation throughput: 2929.9 tokens/s, Running: 256 reqs, Waiting: 475 reqs, GPU KV cache usage: 43.8%, Prefix cache hit rate: 93.3%
+(APIServer pid=316100) INFO 07-15 19:55:26 [loggers.py:273] Engine 000: Avg prompt throughput: 871.9 tokens/s, Avg generation throughput: 2932.8 tokens/s, Running: 256 reqs, Waiting: 363 reqs, GPU KV cache usage: 46.0%, Prefix cache hit rate: 93.4%
+(APIServer pid=316100) INFO 07-15 19:55:36 [loggers.py:273] Engine 000: Avg prompt throughput: 1042.4 tokens/s, Avg generation throughput: 2905.9 tokens/s, Running: 255 reqs, Waiting: 240 reqs, GPU KV cache usage: 45.9%, Prefix cache hit rate: 93.4%
+(APIServer pid=316100) INFO 07-15 19:55:46 [loggers.py:273] Engine 000: Avg prompt throughput: 892.9 tokens/s, Avg generation throughput: 2957.9 tokens/s, Running: 255 reqs, Waiting: 123 reqs, GPU KV cache usage: 46.4%, Prefix cache hit rate: 93.4%
+(APIServer pid=316100) INFO 07-15 19:55:56 [loggers.py:273] Engine 000: Avg prompt throughput: 963.2 tokens/s, Avg generation throughput: 2905.8 tokens/s, Running: 256 reqs, Waiting: 6 reqs, GPU KV cache usage: 48.4%, Prefix cache hit rate: 93.4%
+(APIServer pid=316100) INFO 07-15 19:56:06 [loggers.py:273] Engine 000: Avg prompt throughput: 57.3 tokens/s, Avg generation throughput: 2710.3 tokens/s, Running: 113 reqs, Waiting: 0 reqs, GPU KV cache usage: 29.6%, Prefix cache hit rate: 93.4%
+(APIServer pid=316100) INFO 07-15 19:56:16 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1039.4 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.1%, Prefix cache hit rate: 93.4%
+(APIServer pid=316100) INFO:     127.0.0.1:38272 - "POST /v1/completions HTTP/1.1" 200 OK
+(EngineCore pid=316217) INFO 07-15 19:56:18 [core.py:1214] [shutdown] EngineCore: trigger received signal=SIGTERM
+(APIServer pid=316100) INFO 07-15 19:56:18 [launcher.py:100] [shutdown] API server: shutdown triggered
+(APIServer pid=316100) INFO 07-15 19:56:18 [launcher.py:116] [shutdown] API server: stopping engine client mode=abort timeout=0s
+(EngineCore pid=316217) INFO 07-15 19:56:18 [core.py:1333] [shutdown] EngineCore: start mode=abort timeout=0s
+(EngineCore pid=316217) INFO 07-15 19:56:18 [core.py:1364] [shutdown] EngineCore: request processing complete; starting resource teardown
+(EngineCore pid=316217) INFO 07-15 19:56:18 [core.py:1227] [shutdown] EngineCore: exiting busy loop
+(APIServer pid=316100) INFO 07-15 19:56:18 [core_client.py:655] [shutdown] MPClient: start timeout=0s
+(APIServer pid=316100) INFO 07-15 19:56:18 [core_client.py:657] [shutdown] MPClient: stopping engine manager
+(APIServer pid=316100) WARNING 07-15 19:56:18 [utils.py:626] [shutdown] Process manager: force killing remaining processes count=1
+(APIServer pid=316100) INFO 07-15 19:56:18 [core_client.py:659] [shutdown] MPClient: engine manager stopped
+(APIServer pid=316100) INFO 07-15 19:56:18 [core_client.py:660] [shutdown] MPClient: cleaning up background resources
+(APIServer pid=316100) INFO 07-15 19:56:18 [core_client.py:662] [shutdown] MPClient: complete
+(APIServer pid=316100) INFO 07-15 19:56:18 [launcher.py:125] [shutdown] API server: engine client stopped
+(APIServer pid=316100) INFO 07-15 19:56:18 [launcher.py:128] [shutdown] API server: signalling HTTP server shutdown
+(APIServer pid=316100) INFO 07-15 19:56:18 [launcher.py:149] [shutdown] API server: shutting down FastAPI HTTP server
+(APIServer pid=316100) INFO:     Shutting down
+(APIServer pid=316100) INFO:     Waiting for application shutdown.
+(APIServer pid=316100) INFO:     Application shutdown complete.
+/home/henry/.local/share/uv/python/cpython-3.12.12-linux-x86_64-gnu/lib/python3.12/multiprocessing/resource_tracker.py:279: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
+  warnings.warn('resource_tracker: There appear to be %d '

evals/healing_breadth/keep25_traj/step0250_chat.json.server.log

@@ -0,0 +1,134 @@
+(APIServer pid=317520) INFO 07-15 19:56:43 [api_utils.py:339] 
+(APIServer pid=317520) INFO 07-15 19:56:43 [api_utils.py:339]        █     █     █▄   ▄█
+(APIServer pid=317520) INFO 07-15 19:56:43 [api_utils.py:339]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.25.0
+(APIServer pid=317520) INFO 07-15 19:56:43 [api_utils.py:339]   █▄█▀ █     █     █     █  model   outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0250
+(APIServer pid=317520) INFO 07-15 19:56:43 [api_utils.py:339]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
+(APIServer pid=317520) INFO 07-15 19:56:43 [api_utils.py:339] 
+(APIServer pid=317520) INFO 07-15 19:56:43 [api_utils.py:273] non-default args: {'model_tag': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0250', 'host': '127.0.0.1', 'port': 8377, 'model': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0250', 'max_model_len': 2048, 'enforce_eager': True, 'served_model_name': ['student'], 'gpu_memory_utilization': 0.85}
+(APIServer pid=317520) INFO 07-15 19:56:43 [model.py:619] Resolved architecture: PrunedOlmoeForCausalLM
+(APIServer pid=317520) INFO 07-15 19:56:43 [model.py:1776] Using max model len 2048
+(APIServer pid=317520) INFO 07-15 19:56:43 [vllm.py:1042] Asynchronous scheduling is enabled.
+(APIServer pid=317520) WARNING 07-15 19:56:43 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(APIServer pid=317520) WARNING 07-15 19:56:43 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(APIServer pid=317520) INFO 07-15 19:56:43 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(APIServer pid=317520) INFO 07-15 19:56:43 [vllm.py:1322] Cudagraph is disabled under eager mode
+(APIServer pid=317520) INFO 07-15 19:56:43 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(EngineCore pid=317635) INFO 07-15 19:56:50 [core.py:114] Initializing a V1 LLM engine (v0.25.0) with config: model='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0250', speculative_config=None, tokenizer='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0250', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=student, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.NONE: 0>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, moe_backend='auto', linear_backend='auto')
+(EngineCore pid=317635) INFO 07-15 19:56:50 [parallel_state.py:1607] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.0.15:41309 backend=nccl
+(EngineCore pid=317635) INFO 07-15 19:56:50 [parallel_state.py:1942] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A
+(EngineCore pid=317635) INFO 07-15 19:56:51 [topk_topp_sampler.py:55] Using FlashInfer for top-p & top-k sampling.
+(EngineCore pid=317635) INFO 07-15 19:56:51 [gpu_model_runner.py:5209] Starting to load model outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0250...
+(EngineCore pid=317635) INFO 07-15 19:56:52 [cuda.py:476] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
+(EngineCore pid=317635) INFO 07-15 19:56:52 [flash_attn.py:718] Using FlashAttention version 2
+(EngineCore pid=317635) /home/henry/.cache/glean/megablocks-variable-93a1479bc15b/megablocks/grouped_gemm_util.py:10: UserWarning: Grouped GEMM not available.
+(EngineCore pid=317635)   warnings.warn('Grouped GEMM not available.')
+(EngineCore pid=317635) INFO 07-15 19:56:52 [weight_utils.py:849] Filesystem type for checkpoints: EXT4. Checkpoint size: 3.89 GiB. Available RAM: 106.43 GiB.
+(EngineCore pid=317635) INFO 07-15 19:56:52 [weight_utils.py:872] Auto-prefetch is disabled because the filesystem (EXT4) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.
+(EngineCore pid=317635) 
+(EngineCore pid=317635) 
+(EngineCore pid=317635) 
+(EngineCore pid=317635) 
+(EngineCore pid=317635) INFO 07-15 19:56:54 [default_loader.py:430] Loading weights took 2.49 seconds
+(EngineCore pid=317635) INFO 07-15 19:56:55 [gpu_model_runner.py:5306] Model loading took 3.89 GiB memory and 2.674049 seconds
+(EngineCore pid=317635) INFO 07-15 19:56:56 [gpu_worker.py:538] Available KV cache memory: 15.82 GiB
+(EngineCore pid=317635) INFO 07-15 19:56:56 [kv_cache_utils.py:2146] GPU KV cache size: 129,584 tokens
+(EngineCore pid=317635) INFO 07-15 19:56:56 [kv_cache_utils.py:2147] Maximum concurrency for 2,048 tokens per request: 63.27x
+(EngineCore pid=317635) INFO 07-15 19:56:57 [cutedsl_warmup.py:97] Skipping CuTeDSL warmup because no compile units were requested.
+(EngineCore pid=317635) INFO 07-15 19:56:57 [jit_monitor.py:73] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
+(EngineCore pid=317635) INFO 07-15 19:56:57 [core.py:344] init engine (profile, create kv cache, warmup model) took 2.04 s
+(EngineCore pid=317635) INFO 07-15 19:56:57 [vllm.py:1042] Asynchronous scheduling is enabled.
+(EngineCore pid=317635) WARNING 07-15 19:56:57 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(EngineCore pid=317635) WARNING 07-15 19:56:57 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(EngineCore pid=317635) INFO 07-15 19:56:57 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(EngineCore pid=317635) INFO 07-15 19:56:57 [vllm.py:1322] Cudagraph is disabled under eager mode
+(EngineCore pid=317635) INFO 07-15 19:56:57 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(APIServer pid=317520) INFO 07-15 19:56:57 [api_server.py:612] Supported tasks: ['generate']
+(APIServer pid=317520) WARNING 07-15 19:56:57 [__init__.py:36] SECURITY WARNING: Development endpoints are enabled! This should NOT be used in production!
+(APIServer pid=317520) INFO 07-15 19:56:57 [hf.py:548] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
+(APIServer pid=317520) INFO 07-15 19:56:57 [api_server.py:616] Starting vLLM server on http://127.0.0.1:8377
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:37] Available routes are:
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /openapi.json, Methods: HEAD, GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /docs, Methods: HEAD, GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: HEAD, GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /redoc, Methods: HEAD, GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /load, Methods: GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /version, Methods: GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /health, Methods: GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /metrics, Methods: GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /tokenize, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /detokenize, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /v1/models, Methods: GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /ping, Methods: GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /ping, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /invocations, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /reset_prefix_cache, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /reset_mm_cache, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /reset_encoder_cache, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /pause, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /resume, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /is_paused, Methods: GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /init_weight_transfer_engine, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /start_weight_update, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /update_weights, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /finish_weight_update, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /get_world_size, Methods: GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /collective_rpc, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /server_info, Methods: GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /sleep, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /wake_up, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /is_sleeping, Methods: GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /v1/chat/completions/batch, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /v1/responses, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /v1/completions, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /v1/messages, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /generative_scoring, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /v1/completions/render, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /v1/chat/completions/derender, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /v1/completions/derender, Methods: POST
+(APIServer pid=317520) INFO 07-15 19:56:57 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
+(APIServer pid=317520) INFO:     Started server process [317520]
+(APIServer pid=317520) INFO:     Waiting for application startup.
+(APIServer pid=317520) INFO:     Application startup complete.
+(APIServer pid=317520) INFO:     127.0.0.1:48054 - "GET /health HTTP/1.1" 200 OK
+(EngineCore pid=317635) WARNING 07-15 19:56:59 [jit_monitor.py:129] Triton kernel JIT compilation during inference: _build_route_rows. This causes a latency spike; consider extending warmup to cover this shape/config.
+(APIServer pid=317520) INFO 07-15 19:57:08 [loggers.py:273] Engine 000: Avg prompt throughput: 2065.9 tokens/s, Avg generation throughput: 2243.0 tokens/s, Running: 253 reqs, Waiting: 1056 reqs, GPU KV cache usage: 35.5%, Prefix cache hit rate: 93.1%
+(APIServer pid=317520) INFO 07-15 19:57:18 [loggers.py:273] Engine 000: Avg prompt throughput: 757.9 tokens/s, Avg generation throughput: 2959.8 tokens/s, Running: 253 reqs, Waiting: 959 reqs, GPU KV cache usage: 46.2%, Prefix cache hit rate: 93.2%
+(APIServer pid=317520) INFO 07-15 19:57:28 [loggers.py:273] Engine 000: Avg prompt throughput: 941.0 tokens/s, Avg generation throughput: 2931.4 tokens/s, Running: 256 reqs, Waiting: 840 reqs, GPU KV cache usage: 47.1%, Prefix cache hit rate: 93.2%
+(APIServer pid=317520) INFO 07-15 19:57:38 [loggers.py:273] Engine 000: Avg prompt throughput: 790.5 tokens/s, Avg generation throughput: 2907.5 tokens/s, Running: 254 reqs, Waiting: 736 reqs, GPU KV cache usage: 49.7%, Prefix cache hit rate: 93.3%
+(APIServer pid=317520) INFO 07-15 19:57:48 [loggers.py:273] Engine 000: Avg prompt throughput: 985.7 tokens/s, Avg generation throughput: 2905.8 tokens/s, Running: 256 reqs, Waiting: 611 reqs, GPU KV cache usage: 45.4%, Prefix cache hit rate: 93.3%
+(APIServer pid=317520) INFO 07-15 19:57:58 [loggers.py:273] Engine 000: Avg prompt throughput: 873.5 tokens/s, Avg generation throughput: 2907.7 tokens/s, Running: 256 reqs, Waiting: 503 reqs, GPU KV cache usage: 47.9%, Prefix cache hit rate: 93.3%
+(APIServer pid=317520) INFO 07-15 19:58:08 [loggers.py:273] Engine 000: Avg prompt throughput: 1003.8 tokens/s, Avg generation throughput: 2930.8 tokens/s, Running: 254 reqs, Waiting: 373 reqs, GPU KV cache usage: 45.2%, Prefix cache hit rate: 93.3%
+(APIServer pid=317520) INFO 07-15 19:58:18 [loggers.py:273] Engine 000: Avg prompt throughput: 926.5 tokens/s, Avg generation throughput: 2957.5 tokens/s, Running: 256 reqs, Waiting: 263 reqs, GPU KV cache usage: 48.0%, Prefix cache hit rate: 93.4%
+(APIServer pid=317520) INFO 07-15 19:58:28 [loggers.py:273] Engine 000: Avg prompt throughput: 921.8 tokens/s, Avg generation throughput: 2932.1 tokens/s, Running: 255 reqs, Waiting: 144 reqs, GPU KV cache usage: 47.2%, Prefix cache hit rate: 93.4%
+(APIServer pid=317520) INFO 07-15 19:58:38 [loggers.py:273] Engine 000: Avg prompt throughput: 822.2 tokens/s, Avg generation throughput: 2908.2 tokens/s, Running: 255 reqs, Waiting: 45 reqs, GPU KV cache usage: 50.9%, Prefix cache hit rate: 93.4%
+(APIServer pid=317520) INFO 07-15 19:58:48 [loggers.py:273] Engine 000: Avg prompt throughput: 358.8 tokens/s, Avg generation throughput: 2873.6 tokens/s, Running: 196 reqs, Waiting: 0 reqs, GPU KV cache usage: 47.0%, Prefix cache hit rate: 93.4%
+(APIServer pid=317520) INFO 07-15 19:58:58 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2023.8 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.8%, Prefix cache hit rate: 93.4%
+(APIServer pid=317520) INFO:     127.0.0.1:48058 - "POST /v1/completions HTTP/1.1" 200 OK
+(EngineCore pid=317635) INFO 07-15 19:59:04 [core.py:1214] [shutdown] EngineCore: trigger received signal=SIGTERM
+(APIServer pid=317520) INFO 07-15 19:59:04 [launcher.py:100] [shutdown] API server: shutdown triggered
+(EngineCore pid=317635) INFO 07-15 19:59:04 [core.py:1333] [shutdown] EngineCore: start mode=abort timeout=0s
+(EngineCore pid=317635) INFO 07-15 19:59:04 [core.py:1364] [shutdown] EngineCore: request processing complete; starting resource teardown
+(EngineCore pid=317635) INFO 07-15 19:59:04 [core.py:1227] [shutdown] EngineCore: exiting busy loop
+(APIServer pid=317520) INFO 07-15 19:59:04 [launcher.py:116] [shutdown] API server: stopping engine client mode=abort timeout=0s
+(APIServer pid=317520) INFO 07-15 19:59:04 [core_client.py:655] [shutdown] MPClient: start timeout=0s
+(APIServer pid=317520) INFO 07-15 19:59:04 [core_client.py:657] [shutdown] MPClient: stopping engine manager
+(APIServer pid=317520) WARNING 07-15 19:59:04 [utils.py:626] [shutdown] Process manager: force killing remaining processes count=1
+(APIServer pid=317520) INFO:     Shutting down
+(APIServer pid=317520) INFO 07-15 19:59:04 [core_client.py:659] [shutdown] MPClient: engine manager stopped
+(APIServer pid=317520) INFO 07-15 19:59:04 [core_client.py:660] [shutdown] MPClient: cleaning up background resources
+(APIServer pid=317520) INFO 07-15 19:59:04 [core_client.py:662] [shutdown] MPClient: complete
+(APIServer pid=317520) INFO 07-15 19:59:04 [launcher.py:125] [shutdown] API server: engine client stopped
+(APIServer pid=317520) INFO 07-15 19:59:04 [launcher.py:128] [shutdown] API server: signalling HTTP server shutdown
+(APIServer pid=317520) INFO 07-15 19:59:04 [launcher.py:149] [shutdown] API server: shutting down FastAPI HTTP server
+(APIServer pid=317520) INFO:     Shutting down
+(APIServer pid=317520) INFO:     Waiting for application shutdown.
+(APIServer pid=317520) INFO:     Application shutdown complete.
+/home/henry/.local/share/uv/python/cpython-3.12.12-linux-x86_64-gnu/lib/python3.12/multiprocessing/resource_tracker.py:279: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
+  warnings.warn('resource_tracker: There appear to be %d '

evals/healing_breadth/keep25_traj/step0300_chat.json.server.log

@@ -0,0 +1,133 @@
+(APIServer pid=318952) INFO 07-15 19:59:29 [api_utils.py:339] 
+(APIServer pid=318952) INFO 07-15 19:59:29 [api_utils.py:339]        █     █     █▄   ▄█
+(APIServer pid=318952) INFO 07-15 19:59:29 [api_utils.py:339]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.25.0
+(APIServer pid=318952) INFO 07-15 19:59:29 [api_utils.py:339]   █▄█▀ █     █     █     █  model   outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0300
+(APIServer pid=318952) INFO 07-15 19:59:29 [api_utils.py:339]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
+(APIServer pid=318952) INFO 07-15 19:59:29 [api_utils.py:339] 
+(APIServer pid=318952) INFO 07-15 19:59:29 [api_utils.py:273] non-default args: {'model_tag': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0300', 'host': '127.0.0.1', 'port': 8377, 'model': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0300', 'max_model_len': 2048, 'enforce_eager': True, 'served_model_name': ['student'], 'gpu_memory_utilization': 0.85}
+(APIServer pid=318952) INFO 07-15 19:59:29 [model.py:619] Resolved architecture: PrunedOlmoeForCausalLM
+(APIServer pid=318952) INFO 07-15 19:59:29 [model.py:1776] Using max model len 2048
+(APIServer pid=318952) INFO 07-15 19:59:29 [vllm.py:1042] Asynchronous scheduling is enabled.
+(APIServer pid=318952) WARNING 07-15 19:59:29 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(APIServer pid=318952) WARNING 07-15 19:59:29 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(APIServer pid=318952) INFO 07-15 19:59:29 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(APIServer pid=318952) INFO 07-15 19:59:29 [vllm.py:1322] Cudagraph is disabled under eager mode
+(APIServer pid=318952) INFO 07-15 19:59:29 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(EngineCore pid=319071) INFO 07-15 19:59:36 [core.py:114] Initializing a V1 LLM engine (v0.25.0) with config: model='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0300', speculative_config=None, tokenizer='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0300', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=student, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.NONE: 0>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, moe_backend='auto', linear_backend='auto')
+(EngineCore pid=319071) INFO 07-15 19:59:37 [parallel_state.py:1607] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.0.15:46937 backend=nccl
+(EngineCore pid=319071) INFO 07-15 19:59:37 [parallel_state.py:1942] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A
+(EngineCore pid=319071) INFO 07-15 19:59:38 [topk_topp_sampler.py:55] Using FlashInfer for top-p & top-k sampling.
+(EngineCore pid=319071) INFO 07-15 19:59:38 [gpu_model_runner.py:5209] Starting to load model outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0300...
+(EngineCore pid=319071) INFO 07-15 19:59:38 [cuda.py:476] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
+(EngineCore pid=319071) INFO 07-15 19:59:38 [flash_attn.py:718] Using FlashAttention version 2
+(EngineCore pid=319071) /home/henry/.cache/glean/megablocks-variable-93a1479bc15b/megablocks/grouped_gemm_util.py:10: UserWarning: Grouped GEMM not available.
+(EngineCore pid=319071)   warnings.warn('Grouped GEMM not available.')
+(EngineCore pid=319071) INFO 07-15 19:59:38 [weight_utils.py:849] Filesystem type for checkpoints: EXT4. Checkpoint size: 3.89 GiB. Available RAM: 106.50 GiB.
+(EngineCore pid=319071) INFO 07-15 19:59:38 [weight_utils.py:872] Auto-prefetch is disabled because the filesystem (EXT4) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.
+(EngineCore pid=319071) 
+(EngineCore pid=319071) 
+(EngineCore pid=319071) 
+(EngineCore pid=319071) 
+(EngineCore pid=319071) INFO 07-15 19:59:41 [default_loader.py:430] Loading weights took 2.39 seconds
+(EngineCore pid=319071) INFO 07-15 19:59:41 [gpu_model_runner.py:5306] Model loading took 3.89 GiB memory and 2.573598 seconds
+(EngineCore pid=319071) INFO 07-15 19:59:43 [gpu_worker.py:538] Available KV cache memory: 15.82 GiB
+(EngineCore pid=319071) INFO 07-15 19:59:43 [kv_cache_utils.py:2146] GPU KV cache size: 129,584 tokens
+(EngineCore pid=319071) INFO 07-15 19:59:43 [kv_cache_utils.py:2147] Maximum concurrency for 2,048 tokens per request: 63.27x
+(EngineCore pid=319071) INFO 07-15 19:59:43 [cutedsl_warmup.py:97] Skipping CuTeDSL warmup because no compile units were requested.
+(EngineCore pid=319071) INFO 07-15 19:59:43 [jit_monitor.py:73] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
+(EngineCore pid=319071) INFO 07-15 19:59:43 [core.py:344] init engine (profile, create kv cache, warmup model) took 2.07 s
+(EngineCore pid=319071) INFO 07-15 19:59:43 [vllm.py:1042] Asynchronous scheduling is enabled.
+(EngineCore pid=319071) WARNING 07-15 19:59:43 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(EngineCore pid=319071) WARNING 07-15 19:59:43 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(EngineCore pid=319071) INFO 07-15 19:59:43 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(EngineCore pid=319071) INFO 07-15 19:59:43 [vllm.py:1322] Cudagraph is disabled under eager mode
+(EngineCore pid=319071) INFO 07-15 19:59:43 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(APIServer pid=318952) INFO 07-15 19:59:43 [api_server.py:612] Supported tasks: ['generate']
+(APIServer pid=318952) WARNING 07-15 19:59:44 [__init__.py:36] SECURITY WARNING: Development endpoints are enabled! This should NOT be used in production!
+(APIServer pid=318952) INFO 07-15 19:59:44 [hf.py:548] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
+(APIServer pid=318952) INFO 07-15 19:59:44 [api_server.py:616] Starting vLLM server on http://127.0.0.1:8377
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:37] Available routes are:
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /openapi.json, Methods: GET, HEAD
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /docs, Methods: GET, HEAD
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: GET, HEAD
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /redoc, Methods: GET, HEAD
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /load, Methods: GET
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /version, Methods: GET
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /health, Methods: GET
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /metrics, Methods: GET
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /tokenize, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /detokenize, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /v1/models, Methods: GET
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /ping, Methods: GET
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /ping, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /invocations, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /reset_prefix_cache, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /reset_mm_cache, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /reset_encoder_cache, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /pause, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /resume, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /is_paused, Methods: GET
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /init_weight_transfer_engine, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /start_weight_update, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /update_weights, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /finish_weight_update, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /get_world_size, Methods: GET
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /collective_rpc, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /server_info, Methods: GET
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /sleep, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /wake_up, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /is_sleeping, Methods: GET
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /v1/chat/completions/batch, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /v1/responses, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /v1/completions, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /v1/messages, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /generative_scoring, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /v1/completions/render, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /v1/chat/completions/derender, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /v1/completions/derender, Methods: POST
+(APIServer pid=318952) INFO 07-15 19:59:44 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
+(APIServer pid=318952) INFO:     Started server process [318952]
+(APIServer pid=318952) INFO:     Waiting for application startup.
+(APIServer pid=318952) INFO:     Application startup complete.
+(APIServer pid=318952) INFO:     127.0.0.1:46904 - "GET /health HTTP/1.1" 200 OK
+(EngineCore pid=319071) WARNING 07-15 19:59:46 [jit_monitor.py:129] Triton kernel JIT compilation during inference: _build_route_rows. This causes a latency spike; consider extending warmup to cover this shape/config.
+(APIServer pid=318952) INFO 07-15 19:59:54 [loggers.py:273] Engine 000: Avg prompt throughput: 2104.6 tokens/s, Avg generation throughput: 2218.7 tokens/s, Running: 256 reqs, Waiting: 1051 reqs, GPU KV cache usage: 35.5%, Prefix cache hit rate: 93.1%
+(APIServer pid=318952) INFO 07-15 20:00:04 [loggers.py:273] Engine 000: Avg prompt throughput: 770.5 tokens/s, Avg generation throughput: 2958.6 tokens/s, Running: 254 reqs, Waiting: 952 reqs, GPU KV cache usage: 46.4%, Prefix cache hit rate: 93.2%
+(APIServer pid=318952) INFO 07-15 20:00:14 [loggers.py:273] Engine 000: Avg prompt throughput: 1022.2 tokens/s, Avg generation throughput: 2930.3 tokens/s, Running: 254 reqs, Waiting: 822 reqs, GPU KV cache usage: 45.6%, Prefix cache hit rate: 93.3%
+(APIServer pid=318952) INFO 07-15 20:00:24 [loggers.py:273] Engine 000: Avg prompt throughput: 759.2 tokens/s, Avg generation throughput: 2908.6 tokens/s, Running: 252 reqs, Waiting: 725 reqs, GPU KV cache usage: 50.7%, Prefix cache hit rate: 93.3%
+(APIServer pid=318952) INFO 07-15 20:00:34 [loggers.py:273] Engine 000: Avg prompt throughput: 1117.1 tokens/s, Avg generation throughput: 2903.9 tokens/s, Running: 256 reqs, Waiting: 580 reqs, GPU KV cache usage: 42.9%, Prefix cache hit rate: 93.3%
+(APIServer pid=318952) INFO 07-15 20:00:44 [loggers.py:273] Engine 000: Avg prompt throughput: 910.0 tokens/s, Avg generation throughput: 2932.4 tokens/s, Running: 254 reqs, Waiting: 465 reqs, GPU KV cache usage: 45.3%, Prefix cache hit rate: 93.3%
+(APIServer pid=318952) INFO 07-15 20:00:54 [loggers.py:273] Engine 000: Avg prompt throughput: 976.7 tokens/s, Avg generation throughput: 2931.5 tokens/s, Running: 254 reqs, Waiting: 345 reqs, GPU KV cache usage: 44.7%, Prefix cache hit rate: 93.3%
+(APIServer pid=318952) INFO 07-15 20:01:04 [loggers.py:273] Engine 000: Avg prompt throughput: 980.9 tokens/s, Avg generation throughput: 2931.5 tokens/s, Running: 255 reqs, Waiting: 226 reqs, GPU KV cache usage: 47.0%, Prefix cache hit rate: 93.4%
+(APIServer pid=318952) INFO 07-15 20:01:14 [loggers.py:273] Engine 000: Avg prompt throughput: 970.9 tokens/s, Avg generation throughput: 2931.3 tokens/s, Running: 256 reqs, Waiting: 103 reqs, GPU KV cache usage: 46.7%, Prefix cache hit rate: 93.4%
+(APIServer pid=318952) INFO 07-15 20:01:24 [loggers.py:273] Engine 000: Avg prompt throughput: 831.4 tokens/s, Avg generation throughput: 2901.6 tokens/s, Running: 245 reqs, Waiting: 0 reqs, GPU KV cache usage: 47.8%, Prefix cache hit rate: 93.4%
+(APIServer pid=318952) INFO 07-15 20:01:34 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2619.3 tokens/s, Running: 101 reqs, Waiting: 0 reqs, GPU KV cache usage: 29.6%, Prefix cache hit rate: 93.4%
+(APIServer pid=318952) INFO 07-15 20:01:44 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 855.3 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.3%, Prefix cache hit rate: 93.4%
+(APIServer pid=318952) INFO:     127.0.0.1:46914 - "POST /v1/completions HTTP/1.1" 200 OK
+(EngineCore pid=319071) INFO 07-15 20:01:45 [core.py:1214] [shutdown] EngineCore: trigger received signal=SIGTERM
+(APIServer pid=318952) INFO 07-15 20:01:45 [launcher.py:100] [shutdown] API server: shutdown triggered
+(APIServer pid=318952) INFO 07-15 20:01:45 [launcher.py:116] [shutdown] API server: stopping engine client mode=abort timeout=0s
+(EngineCore pid=319071) INFO 07-15 20:01:45 [core.py:1333] [shutdown] EngineCore: start mode=abort timeout=0s
+(EngineCore pid=319071) INFO 07-15 20:01:45 [core.py:1364] [shutdown] EngineCore: request processing complete; starting resource teardown
+(EngineCore pid=319071) INFO 07-15 20:01:45 [core.py:1227] [shutdown] EngineCore: exiting busy loop
+(APIServer pid=318952) INFO 07-15 20:01:45 [core_client.py:655] [shutdown] MPClient: start timeout=0s
+(APIServer pid=318952) INFO 07-15 20:01:45 [core_client.py:657] [shutdown] MPClient: stopping engine manager
+(APIServer pid=318952) WARNING 07-15 20:01:45 [utils.py:626] [shutdown] Process manager: force killing remaining processes count=1
+(APIServer pid=318952) INFO 07-15 20:01:45 [core_client.py:659] [shutdown] MPClient: engine manager stopped
+(APIServer pid=318952) INFO 07-15 20:01:45 [core_client.py:660] [shutdown] MPClient: cleaning up background resources
+(APIServer pid=318952) INFO 07-15 20:01:45 [core_client.py:662] [shutdown] MPClient: complete
+(APIServer pid=318952) INFO 07-15 20:01:45 [launcher.py:125] [shutdown] API server: engine client stopped
+(APIServer pid=318952) INFO 07-15 20:01:45 [launcher.py:128] [shutdown] API server: signalling HTTP server shutdown
+(APIServer pid=318952) INFO 07-15 20:01:45 [launcher.py:149] [shutdown] API server: shutting down FastAPI HTTP server
+(APIServer pid=318952) INFO:     Shutting down
+(APIServer pid=318952) INFO:     Waiting for application shutdown.
+(APIServer pid=318952) INFO:     Application shutdown complete.
+/home/henry/.local/share/uv/python/cpython-3.12.12-linux-x86_64-gnu/lib/python3.12/multiprocessing/resource_tracker.py:279: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
+  warnings.warn('resource_tracker: There appear to be %d '

evals/healing_breadth/keep25_traj/step0350_chat.json.server.log

@@ -0,0 +1,133 @@
+(APIServer pid=315350) INFO 07-15 19:52:00 [api_utils.py:339] 
+(APIServer pid=315350) INFO 07-15 19:52:00 [api_utils.py:339]        █     █     █▄   ▄█
+(APIServer pid=315350) INFO 07-15 19:52:00 [api_utils.py:339]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.25.0
+(APIServer pid=315350) INFO 07-15 19:52:00 [api_utils.py:339]   █▄█▀ █     █     █     █  model   outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0350
+(APIServer pid=315350) INFO 07-15 19:52:00 [api_utils.py:339]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
+(APIServer pid=315350) INFO 07-15 19:52:00 [api_utils.py:339] 
+(APIServer pid=315350) INFO 07-15 19:52:00 [api_utils.py:273] non-default args: {'model_tag': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0350', 'host': '127.0.0.1', 'port': 8378, 'model': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0350', 'max_model_len': 2048, 'enforce_eager': True, 'served_model_name': ['student'], 'gpu_memory_utilization': 0.85}
+(APIServer pid=315350) INFO 07-15 19:52:00 [model.py:619] Resolved architecture: PrunedOlmoeForCausalLM
+(APIServer pid=315350) INFO 07-15 19:52:00 [model.py:1776] Using max model len 2048
+(APIServer pid=315350) INFO 07-15 19:52:00 [vllm.py:1042] Asynchronous scheduling is enabled.
+(APIServer pid=315350) WARNING 07-15 19:52:00 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(APIServer pid=315350) WARNING 07-15 19:52:00 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(APIServer pid=315350) INFO 07-15 19:52:00 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(APIServer pid=315350) INFO 07-15 19:52:00 [vllm.py:1322] Cudagraph is disabled under eager mode
+(APIServer pid=315350) INFO 07-15 19:52:00 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(EngineCore pid=315465) INFO 07-15 19:52:07 [core.py:114] Initializing a V1 LLM engine (v0.25.0) with config: model='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0350', speculative_config=None, tokenizer='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0350', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=student, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.NONE: 0>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, moe_backend='auto', linear_backend='auto')
+(EngineCore pid=315465) INFO 07-15 19:52:08 [parallel_state.py:1607] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.0.15:37247 backend=nccl
+(EngineCore pid=315465) INFO 07-15 19:52:08 [parallel_state.py:1942] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A
+(EngineCore pid=315465) INFO 07-15 19:52:09 [topk_topp_sampler.py:55] Using FlashInfer for top-p & top-k sampling.
+(EngineCore pid=315465) INFO 07-15 19:52:09 [gpu_model_runner.py:5209] Starting to load model outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0350...
+(EngineCore pid=315465) INFO 07-15 19:52:09 [cuda.py:476] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
+(EngineCore pid=315465) INFO 07-15 19:52:09 [flash_attn.py:718] Using FlashAttention version 2
+(EngineCore pid=315465) /home/henry/.cache/glean/megablocks-variable-93a1479bc15b/megablocks/grouped_gemm_util.py:10: UserWarning: Grouped GEMM not available.
+(EngineCore pid=315465)   warnings.warn('Grouped GEMM not available.')
+(EngineCore pid=315465) INFO 07-15 19:52:09 [weight_utils.py:849] Filesystem type for checkpoints: EXT4. Checkpoint size: 3.89 GiB. Available RAM: 106.62 GiB.
+(EngineCore pid=315465) INFO 07-15 19:52:09 [weight_utils.py:872] Auto-prefetch is disabled because the filesystem (EXT4) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.
+(EngineCore pid=315465) 
+(EngineCore pid=315465) 
+(EngineCore pid=315465) 
+(EngineCore pid=315465) 
+(EngineCore pid=315465) INFO 07-15 19:52:12 [default_loader.py:430] Loading weights took 2.51 seconds
+(EngineCore pid=315465) INFO 07-15 19:52:12 [gpu_model_runner.py:5306] Model loading took 3.89 GiB memory and 2.690210 seconds
+(EngineCore pid=315465) INFO 07-15 19:52:14 [gpu_worker.py:538] Available KV cache memory: 15.82 GiB
+(EngineCore pid=315465) INFO 07-15 19:52:14 [kv_cache_utils.py:2146] GPU KV cache size: 129,584 tokens
+(EngineCore pid=315465) INFO 07-15 19:52:14 [kv_cache_utils.py:2147] Maximum concurrency for 2,048 tokens per request: 63.27x
+(EngineCore pid=315465) INFO 07-15 19:52:14 [cutedsl_warmup.py:97] Skipping CuTeDSL warmup because no compile units were requested.
+(EngineCore pid=315465) INFO 07-15 19:52:14 [jit_monitor.py:73] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
+(EngineCore pid=315465) INFO 07-15 19:52:15 [core.py:344] init engine (profile, create kv cache, warmup model) took 2.08 s
+(EngineCore pid=315465) INFO 07-15 19:52:15 [vllm.py:1042] Asynchronous scheduling is enabled.
+(EngineCore pid=315465) WARNING 07-15 19:52:15 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(EngineCore pid=315465) WARNING 07-15 19:52:15 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(EngineCore pid=315465) INFO 07-15 19:52:15 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(EngineCore pid=315465) INFO 07-15 19:52:15 [vllm.py:1322] Cudagraph is disabled under eager mode
+(EngineCore pid=315465) INFO 07-15 19:52:15 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(APIServer pid=315350) INFO 07-15 19:52:15 [api_server.py:612] Supported tasks: ['generate']
+(APIServer pid=315350) WARNING 07-15 19:52:15 [__init__.py:36] SECURITY WARNING: Development endpoints are enabled! This should NOT be used in production!
+(APIServer pid=315350) INFO 07-15 19:52:15 [hf.py:548] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
+(APIServer pid=315350) INFO 07-15 19:52:15 [api_server.py:616] Starting vLLM server on http://127.0.0.1:8378
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:37] Available routes are:
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /openapi.json, Methods: HEAD, GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /docs, Methods: HEAD, GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: HEAD, GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /redoc, Methods: HEAD, GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /load, Methods: GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /version, Methods: GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /health, Methods: GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /metrics, Methods: GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /tokenize, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /detokenize, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /v1/models, Methods: GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /ping, Methods: GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /ping, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /invocations, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /reset_prefix_cache, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /reset_mm_cache, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /reset_encoder_cache, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /pause, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /resume, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /is_paused, Methods: GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /init_weight_transfer_engine, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /start_weight_update, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /update_weights, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /finish_weight_update, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /get_world_size, Methods: GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /collective_rpc, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /server_info, Methods: GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /sleep, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /wake_up, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /is_sleeping, Methods: GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /v1/chat/completions/batch, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /v1/responses, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /v1/completions, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /v1/messages, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /generative_scoring, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /v1/completions/render, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /v1/chat/completions/derender, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /v1/completions/derender, Methods: POST
+(APIServer pid=315350) INFO 07-15 19:52:15 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
+(APIServer pid=315350) INFO:     Started server process [315350]
+(APIServer pid=315350) INFO:     Waiting for application startup.
+(APIServer pid=315350) INFO:     Application startup complete.
+(APIServer pid=315350) INFO:     127.0.0.1:48888 - "GET /health HTTP/1.1" 200 OK
+(EngineCore pid=315465) WARNING 07-15 19:52:17 [jit_monitor.py:129] Triton kernel JIT compilation during inference: _build_route_rows. This causes a latency spike; consider extending warmup to cover this shape/config.
+(APIServer pid=315350) INFO 07-15 19:52:25 [loggers.py:273] Engine 000: Avg prompt throughput: 2045.0 tokens/s, Avg generation throughput: 2238.3 tokens/s, Running: 256 reqs, Waiting: 1059 reqs, GPU KV cache usage: 36.1%, Prefix cache hit rate: 93.1%
+(APIServer pid=315350) INFO 07-15 19:52:35 [loggers.py:273] Engine 000: Avg prompt throughput: 739.9 tokens/s, Avg generation throughput: 2934.3 tokens/s, Running: 256 reqs, Waiting: 963 reqs, GPU KV cache usage: 47.5%, Prefix cache hit rate: 93.2%
+(APIServer pid=315350) INFO 07-15 19:52:45 [loggers.py:273] Engine 000: Avg prompt throughput: 998.7 tokens/s, Avg generation throughput: 2931.0 tokens/s, Running: 255 reqs, Waiting: 838 reqs, GPU KV cache usage: 46.3%, Prefix cache hit rate: 93.2%
+(APIServer pid=315350) INFO 07-15 19:52:55 [loggers.py:273] Engine 000: Avg prompt throughput: 790.2 tokens/s, Avg generation throughput: 2907.4 tokens/s, Running: 256 reqs, Waiting: 734 reqs, GPU KV cache usage: 49.0%, Prefix cache hit rate: 93.3%
+(APIServer pid=315350) INFO 07-15 19:53:05 [loggers.py:273] Engine 000: Avg prompt throughput: 985.0 tokens/s, Avg generation throughput: 2931.2 tokens/s, Running: 256 reqs, Waiting: 607 reqs, GPU KV cache usage: 44.2%, Prefix cache hit rate: 93.3%
+(APIServer pid=315350) INFO 07-15 19:53:15 [loggers.py:273] Engine 000: Avg prompt throughput: 973.3 tokens/s, Avg generation throughput: 2932.0 tokens/s, Running: 254 reqs, Waiting: 488 reqs, GPU KV cache usage: 45.6%, Prefix cache hit rate: 93.3%
+(APIServer pid=315350) INFO 07-15 19:53:25 [loggers.py:273] Engine 000: Avg prompt throughput: 1003.8 tokens/s, Avg generation throughput: 2930.7 tokens/s, Running: 255 reqs, Waiting: 359 reqs, GPU KV cache usage: 44.2%, Prefix cache hit rate: 93.3%
+(APIServer pid=315350) INFO 07-15 19:53:35 [loggers.py:273] Engine 000: Avg prompt throughput: 943.6 tokens/s, Avg generation throughput: 2932.5 tokens/s, Running: 254 reqs, Waiting: 244 reqs, GPU KV cache usage: 45.7%, Prefix cache hit rate: 93.4%
+(APIServer pid=315350) INFO 07-15 19:53:45 [loggers.py:273] Engine 000: Avg prompt throughput: 847.8 tokens/s, Avg generation throughput: 2932.8 tokens/s, Running: 253 reqs, Waiting: 135 reqs, GPU KV cache usage: 47.0%, Prefix cache hit rate: 93.4%
+(APIServer pid=315350) INFO 07-15 19:53:55 [loggers.py:273] Engine 000: Avg prompt throughput: 1031.3 tokens/s, Avg generation throughput: 2931.6 tokens/s, Running: 255 reqs, Waiting: 12 reqs, GPU KV cache usage: 47.2%, Prefix cache hit rate: 93.4%
+(APIServer pid=315350) INFO 07-15 19:54:05 [loggers.py:273] Engine 000: Avg prompt throughput: 85.6 tokens/s, Avg generation throughput: 2796.3 tokens/s, Running: 143 reqs, Waiting: 0 reqs, GPU KV cache usage: 35.6%, Prefix cache hit rate: 93.4%
+(APIServer pid=315350) INFO 07-15 19:54:15 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1300.1 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%, Prefix cache hit rate: 93.4%
+(APIServer pid=315350) INFO:     127.0.0.1:48896 - "POST /v1/completions HTTP/1.1" 200 OK
+(EngineCore pid=315465) INFO 07-15 19:54:18 [core.py:1214] [shutdown] EngineCore: trigger received signal=SIGTERM
+(APIServer pid=315350) INFO 07-15 19:54:18 [launcher.py:100] [shutdown] API server: shutdown triggered
+(APIServer pid=315350) INFO 07-15 19:54:18 [launcher.py:116] [shutdown] API server: stopping engine client mode=abort timeout=0s
+(EngineCore pid=315465) INFO 07-15 19:54:18 [core.py:1333] [shutdown] EngineCore: start mode=abort timeout=0s
+(EngineCore pid=315465) INFO 07-15 19:54:18 [core.py:1364] [shutdown] EngineCore: request processing complete; starting resource teardown
+(EngineCore pid=315465) INFO 07-15 19:54:18 [core.py:1227] [shutdown] EngineCore: exiting busy loop
+(APIServer pid=315350) INFO 07-15 19:54:18 [core_client.py:655] [shutdown] MPClient: start timeout=0s
+(APIServer pid=315350) INFO 07-15 19:54:18 [core_client.py:657] [shutdown] MPClient: stopping engine manager
+(APIServer pid=315350) WARNING 07-15 19:54:18 [utils.py:626] [shutdown] Process manager: force killing remaining processes count=1
+(APIServer pid=315350) INFO 07-15 19:54:18 [core_client.py:659] [shutdown] MPClient: engine manager stopped
+(APIServer pid=315350) INFO 07-15 19:54:18 [core_client.py:660] [shutdown] MPClient: cleaning up background resources
+(APIServer pid=315350) INFO 07-15 19:54:18 [core_client.py:662] [shutdown] MPClient: complete
+(APIServer pid=315350) INFO 07-15 19:54:18 [launcher.py:125] [shutdown] API server: engine client stopped
+(APIServer pid=315350) INFO 07-15 19:54:18 [launcher.py:128] [shutdown] API server: signalling HTTP server shutdown
+(APIServer pid=315350) INFO 07-15 19:54:18 [launcher.py:149] [shutdown] API server: shutting down FastAPI HTTP server
+(APIServer pid=315350) INFO:     Shutting down
+(APIServer pid=315350) INFO:     Waiting for application shutdown.
+(APIServer pid=315350) INFO:     Application shutdown complete.
+/home/henry/.local/share/uv/python/cpython-3.12.12-linux-x86_64-gnu/lib/python3.12/multiprocessing/resource_tracker.py:279: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
+  warnings.warn('resource_tracker: There appear to be %d '

evals/healing_breadth/keep25_traj/step0400_chat.json.server.log

@@ -0,0 +1,134 @@
+(APIServer pid=316761) INFO 07-15 19:54:43 [api_utils.py:339] 
+(APIServer pid=316761) INFO 07-15 19:54:43 [api_utils.py:339]        █     █     █▄   ▄█
+(APIServer pid=316761) INFO 07-15 19:54:43 [api_utils.py:339]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.25.0
+(APIServer pid=316761) INFO 07-15 19:54:43 [api_utils.py:339]   █▄█▀ █     █     █     █  model   outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0400
+(APIServer pid=316761) INFO 07-15 19:54:43 [api_utils.py:339]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
+(APIServer pid=316761) INFO 07-15 19:54:43 [api_utils.py:339] 
+(APIServer pid=316761) INFO 07-15 19:54:43 [api_utils.py:273] non-default args: {'model_tag': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0400', 'host': '127.0.0.1', 'port': 8378, 'model': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0400', 'max_model_len': 2048, 'enforce_eager': True, 'served_model_name': ['student'], 'gpu_memory_utilization': 0.85}
+(APIServer pid=316761) INFO 07-15 19:54:43 [model.py:619] Resolved architecture: PrunedOlmoeForCausalLM
+(APIServer pid=316761) INFO 07-15 19:54:43 [model.py:1776] Using max model len 2048
+(APIServer pid=316761) INFO 07-15 19:54:43 [vllm.py:1042] Asynchronous scheduling is enabled.
+(APIServer pid=316761) WARNING 07-15 19:54:43 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(APIServer pid=316761) WARNING 07-15 19:54:43 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(APIServer pid=316761) INFO 07-15 19:54:43 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(APIServer pid=316761) INFO 07-15 19:54:43 [vllm.py:1322] Cudagraph is disabled under eager mode
+(APIServer pid=316761) INFO 07-15 19:54:43 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(EngineCore pid=316879) INFO 07-15 19:54:50 [core.py:114] Initializing a V1 LLM engine (v0.25.0) with config: model='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0400', speculative_config=None, tokenizer='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0400', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=student, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.NONE: 0>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, moe_backend='auto', linear_backend='auto')
+(EngineCore pid=316879) INFO 07-15 19:54:51 [parallel_state.py:1607] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.0.15:59755 backend=nccl
+(EngineCore pid=316879) INFO 07-15 19:54:51 [parallel_state.py:1942] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A
+(EngineCore pid=316879) INFO 07-15 19:54:52 [topk_topp_sampler.py:55] Using FlashInfer for top-p & top-k sampling.
+(EngineCore pid=316879) INFO 07-15 19:54:52 [gpu_model_runner.py:5209] Starting to load model outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0400...
+(EngineCore pid=316879) INFO 07-15 19:54:52 [cuda.py:476] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
+(EngineCore pid=316879) INFO 07-15 19:54:52 [flash_attn.py:718] Using FlashAttention version 2
+(EngineCore pid=316879) /home/henry/.cache/glean/megablocks-variable-93a1479bc15b/megablocks/grouped_gemm_util.py:10: UserWarning: Grouped GEMM not available.
+(EngineCore pid=316879)   warnings.warn('Grouped GEMM not available.')
+(EngineCore pid=316879) INFO 07-15 19:54:52 [weight_utils.py:849] Filesystem type for checkpoints: EXT4. Checkpoint size: 3.89 GiB. Available RAM: 106.60 GiB.
+(EngineCore pid=316879) INFO 07-15 19:54:52 [weight_utils.py:872] Auto-prefetch is disabled because the filesystem (EXT4) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.
+(EngineCore pid=316879) 
+(EngineCore pid=316879) 
+(EngineCore pid=316879) 
+(EngineCore pid=316879) 
+(EngineCore pid=316879) INFO 07-15 19:54:55 [default_loader.py:430] Loading weights took 2.62 seconds
+(EngineCore pid=316879) INFO 07-15 19:54:55 [gpu_model_runner.py:5306] Model loading took 3.89 GiB memory and 2.802614 seconds
+(EngineCore pid=316879) INFO 07-15 19:54:57 [gpu_worker.py:538] Available KV cache memory: 15.83 GiB
+(EngineCore pid=316879) INFO 07-15 19:54:57 [kv_cache_utils.py:2146] GPU KV cache size: 129,648 tokens
+(EngineCore pid=316879) INFO 07-15 19:54:57 [kv_cache_utils.py:2147] Maximum concurrency for 2,048 tokens per request: 63.30x
+(EngineCore pid=316879) INFO 07-15 19:54:57 [cutedsl_warmup.py:97] Skipping CuTeDSL warmup because no compile units were requested.
+(EngineCore pid=316879) INFO 07-15 19:54:57 [jit_monitor.py:73] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
+(EngineCore pid=316879) INFO 07-15 19:54:57 [core.py:344] init engine (profile, create kv cache, warmup model) took 2.03 s
+(EngineCore pid=316879) INFO 07-15 19:54:57 [vllm.py:1042] Asynchronous scheduling is enabled.
+(EngineCore pid=316879) WARNING 07-15 19:54:57 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(EngineCore pid=316879) WARNING 07-15 19:54:57 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(EngineCore pid=316879) INFO 07-15 19:54:57 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(EngineCore pid=316879) INFO 07-15 19:54:57 [vllm.py:1322] Cudagraph is disabled under eager mode
+(EngineCore pid=316879) INFO 07-15 19:54:57 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(APIServer pid=316761) INFO 07-15 19:54:57 [api_server.py:612] Supported tasks: ['generate']
+(APIServer pid=316761) WARNING 07-15 19:54:57 [__init__.py:36] SECURITY WARNING: Development endpoints are enabled! This should NOT be used in production!
+(APIServer pid=316761) INFO 07-15 19:54:58 [hf.py:548] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
+(APIServer pid=316761) INFO 07-15 19:54:58 [api_server.py:616] Starting vLLM server on http://127.0.0.1:8378
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:37] Available routes are:
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /openapi.json, Methods: HEAD, GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /docs, Methods: HEAD, GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: HEAD, GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /redoc, Methods: HEAD, GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /load, Methods: GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /version, Methods: GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /health, Methods: GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /metrics, Methods: GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /tokenize, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /detokenize, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /v1/models, Methods: GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /ping, Methods: GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /ping, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /invocations, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /reset_prefix_cache, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /reset_mm_cache, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /reset_encoder_cache, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /pause, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /resume, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /is_paused, Methods: GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /init_weight_transfer_engine, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /start_weight_update, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /update_weights, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /finish_weight_update, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /get_world_size, Methods: GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /collective_rpc, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /server_info, Methods: GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /sleep, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /wake_up, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /is_sleeping, Methods: GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /v1/chat/completions/batch, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /v1/responses, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /v1/completions, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /v1/messages, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /generative_scoring, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /v1/completions/render, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /v1/chat/completions/derender, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /v1/completions/derender, Methods: POST
+(APIServer pid=316761) INFO 07-15 19:54:58 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
+(APIServer pid=316761) INFO:     Started server process [316761]
+(APIServer pid=316761) INFO:     Waiting for application startup.
+(APIServer pid=316761) INFO:     Application startup complete.
+(APIServer pid=316761) INFO:     127.0.0.1:54586 - "GET /health HTTP/1.1" 200 OK
+(EngineCore pid=316879) WARNING 07-15 19:55:00 [jit_monitor.py:129] Triton kernel JIT compilation during inference: _build_route_rows. This causes a latency spike; consider extending warmup to cover this shape/config.
+(APIServer pid=316761) INFO 07-15 19:55:08 [loggers.py:273] Engine 000: Avg prompt throughput: 2054.3 tokens/s, Avg generation throughput: 2241.8 tokens/s, Running: 256 reqs, Waiting: 1058 reqs, GPU KV cache usage: 36.0%, Prefix cache hit rate: 93.1%
+(APIServer pid=316761) INFO 07-15 19:55:18 [loggers.py:273] Engine 000: Avg prompt throughput: 768.7 tokens/s, Avg generation throughput: 2959.2 tokens/s, Running: 253 reqs, Waiting: 959 reqs, GPU KV cache usage: 46.1%, Prefix cache hit rate: 93.2%
+(APIServer pid=316761) INFO 07-15 19:55:28 [loggers.py:273] Engine 000: Avg prompt throughput: 963.6 tokens/s, Avg generation throughput: 2931.5 tokens/s, Running: 254 reqs, Waiting: 838 reqs, GPU KV cache usage: 47.1%, Prefix cache hit rate: 93.2%
+(APIServer pid=316761) INFO 07-15 19:55:38 [loggers.py:273] Engine 000: Avg prompt throughput: 879.1 tokens/s, Avg generation throughput: 2881.1 tokens/s, Running: 256 reqs, Waiting: 722 reqs, GPU KV cache usage: 47.7%, Prefix cache hit rate: 93.3%
+(APIServer pid=316761) INFO 07-15 19:55:48 [loggers.py:273] Engine 000: Avg prompt throughput: 879.4 tokens/s, Avg generation throughput: 2932.5 tokens/s, Running: 255 reqs, Waiting: 609 reqs, GPU KV cache usage: 46.2%, Prefix cache hit rate: 93.3%
+(APIServer pid=316761) INFO 07-15 19:55:58 [loggers.py:273] Engine 000: Avg prompt throughput: 1038.5 tokens/s, Avg generation throughput: 2905.2 tokens/s, Running: 255 reqs, Waiting: 481 reqs, GPU KV cache usage: 45.3%, Prefix cache hit rate: 93.3%
+(APIServer pid=316761) INFO 07-15 19:56:08 [loggers.py:273] Engine 000: Avg prompt throughput: 994.3 tokens/s, Avg generation throughput: 2931.2 tokens/s, Running: 256 reqs, Waiting: 355 reqs, GPU KV cache usage: 44.3%, Prefix cache hit rate: 93.3%
+(APIServer pid=316761) INFO 07-15 19:56:18 [loggers.py:273] Engine 000: Avg prompt throughput: 968.9 tokens/s, Avg generation throughput: 2931.8 tokens/s, Running: 256 reqs, Waiting: 238 reqs, GPU KV cache usage: 46.1%, Prefix cache hit rate: 93.4%
+(APIServer pid=316761) INFO 07-15 19:56:28 [loggers.py:273] Engine 000: Avg prompt throughput: 892.7 tokens/s, Avg generation throughput: 2957.7 tokens/s, Running: 254 reqs, Waiting: 123 reqs, GPU KV cache usage: 46.5%, Prefix cache hit rate: 93.4%
+(APIServer pid=316761) INFO 07-15 19:56:38 [loggers.py:273] Engine 000: Avg prompt throughput: 844.6 tokens/s, Avg generation throughput: 2908.1 tokens/s, Running: 255 reqs, Waiting: 21 reqs, GPU KV cache usage: 50.8%, Prefix cache hit rate: 93.4%
+(APIServer pid=316761) INFO 07-15 19:56:48 [loggers.py:273] Engine 000: Avg prompt throughput: 162.9 tokens/s, Avg generation throughput: 2771.2 tokens/s, Running: 143 reqs, Waiting: 0 reqs, GPU KV cache usage: 36.8%, Prefix cache hit rate: 93.4%
+(APIServer pid=316761) INFO 07-15 19:56:58 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1421.8 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 93.4%
+(APIServer pid=316761) INFO:     127.0.0.1:54590 - "POST /v1/completions HTTP/1.1" 200 OK
+(EngineCore pid=316879) INFO 07-15 19:57:02 [core.py:1214] [shutdown] EngineCore: trigger received signal=SIGTERM
+(APIServer pid=316761) INFO 07-15 19:57:02 [launcher.py:100] [shutdown] API server: shutdown triggered
+(APIServer pid=316761) INFO 07-15 19:57:02 [launcher.py:116] [shutdown] API server: stopping engine client mode=abort timeout=0s
+(EngineCore pid=316879) INFO 07-15 19:57:02 [core.py:1333] [shutdown] EngineCore: start mode=abort timeout=0s
+(EngineCore pid=316879) INFO 07-15 19:57:02 [core.py:1364] [shutdown] EngineCore: request processing complete; starting resource teardown
+(EngineCore pid=316879) INFO 07-15 19:57:02 [core.py:1227] [shutdown] EngineCore: exiting busy loop
+(APIServer pid=316761) INFO 07-15 19:57:02 [core_client.py:655] [shutdown] MPClient: start timeout=0s
+(APIServer pid=316761) INFO 07-15 19:57:02 [core_client.py:657] [shutdown] MPClient: stopping engine manager
+(APIServer pid=316761) WARNING 07-15 19:57:02 [utils.py:626] [shutdown] Process manager: force killing remaining processes count=1
+(APIServer pid=316761) INFO:     Shutting down
+(APIServer pid=316761) INFO 07-15 19:57:02 [core_client.py:659] [shutdown] MPClient: engine manager stopped
+(APIServer pid=316761) INFO 07-15 19:57:02 [core_client.py:660] [shutdown] MPClient: cleaning up background resources
+(APIServer pid=316761) INFO 07-15 19:57:02 [core_client.py:662] [shutdown] MPClient: complete
+(APIServer pid=316761) INFO 07-15 19:57:02 [launcher.py:125] [shutdown] API server: engine client stopped
+(APIServer pid=316761) INFO 07-15 19:57:02 [launcher.py:128] [shutdown] API server: signalling HTTP server shutdown
+(APIServer pid=316761) INFO 07-15 19:57:02 [launcher.py:149] [shutdown] API server: shutting down FastAPI HTTP server
+(APIServer pid=316761) INFO:     Shutting down
+(APIServer pid=316761) INFO:     Waiting for application shutdown.
+(APIServer pid=316761) INFO:     Application shutdown complete.
+/home/henry/.local/share/uv/python/cpython-3.12.12-linux-x86_64-gnu/lib/python3.12/multiprocessing/resource_tracker.py:279: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
+  warnings.warn('resource_tracker: There appear to be %d '

evals/healing_breadth/keep25_traj/step0450_chat.json.server.log

@@ -0,0 +1,133 @@
+(APIServer pid=318177) INFO 07-15 19:57:26 [api_utils.py:339] 
+(APIServer pid=318177) INFO 07-15 19:57:26 [api_utils.py:339]        █     █     █▄   ▄█
+(APIServer pid=318177) INFO 07-15 19:57:26 [api_utils.py:339]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.25.0
+(APIServer pid=318177) INFO 07-15 19:57:26 [api_utils.py:339]   █▄█▀ █     █     █     █  model   outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0450
+(APIServer pid=318177) INFO 07-15 19:57:26 [api_utils.py:339]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
+(APIServer pid=318177) INFO 07-15 19:57:26 [api_utils.py:339] 
+(APIServer pid=318177) INFO 07-15 19:57:26 [api_utils.py:273] non-default args: {'model_tag': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0450', 'host': '127.0.0.1', 'port': 8378, 'model': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0450', 'max_model_len': 2048, 'enforce_eager': True, 'served_model_name': ['student'], 'gpu_memory_utilization': 0.85}
+(APIServer pid=318177) INFO 07-15 19:57:26 [model.py:619] Resolved architecture: PrunedOlmoeForCausalLM
+(APIServer pid=318177) INFO 07-15 19:57:26 [model.py:1776] Using max model len 2048
+(APIServer pid=318177) INFO 07-15 19:57:26 [vllm.py:1042] Asynchronous scheduling is enabled.
+(APIServer pid=318177) WARNING 07-15 19:57:26 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(APIServer pid=318177) WARNING 07-15 19:57:26 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(APIServer pid=318177) INFO 07-15 19:57:26 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(APIServer pid=318177) INFO 07-15 19:57:26 [vllm.py:1322] Cudagraph is disabled under eager mode
+(APIServer pid=318177) INFO 07-15 19:57:26 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(EngineCore pid=318295) INFO 07-15 19:57:33 [core.py:114] Initializing a V1 LLM engine (v0.25.0) with config: model='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0450', speculative_config=None, tokenizer='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0450', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=student, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.NONE: 0>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, moe_backend='auto', linear_backend='auto')
+(EngineCore pid=318295) INFO 07-15 19:57:34 [parallel_state.py:1607] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.0.15:53145 backend=nccl
+(EngineCore pid=318295) INFO 07-15 19:57:34 [parallel_state.py:1942] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A
+(EngineCore pid=318295) INFO 07-15 19:57:35 [topk_topp_sampler.py:55] Using FlashInfer for top-p & top-k sampling.
+(EngineCore pid=318295) INFO 07-15 19:57:35 [gpu_model_runner.py:5209] Starting to load model outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0450...
+(EngineCore pid=318295) INFO 07-15 19:57:35 [cuda.py:476] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
+(EngineCore pid=318295) INFO 07-15 19:57:35 [flash_attn.py:718] Using FlashAttention version 2
+(EngineCore pid=318295) /home/henry/.cache/glean/megablocks-variable-93a1479bc15b/megablocks/grouped_gemm_util.py:10: UserWarning: Grouped GEMM not available.
+(EngineCore pid=318295)   warnings.warn('Grouped GEMM not available.')
+(EngineCore pid=318295) INFO 07-15 19:57:35 [weight_utils.py:849] Filesystem type for checkpoints: EXT4. Checkpoint size: 3.89 GiB. Available RAM: 106.47 GiB.
+(EngineCore pid=318295) INFO 07-15 19:57:35 [weight_utils.py:872] Auto-prefetch is disabled because the filesystem (EXT4) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.
+(EngineCore pid=318295) 
+(EngineCore pid=318295) 
+(EngineCore pid=318295) 
+(EngineCore pid=318295) 
+(EngineCore pid=318295) INFO 07-15 19:57:38 [default_loader.py:430] Loading weights took 2.53 seconds
+(EngineCore pid=318295) INFO 07-15 19:57:38 [gpu_model_runner.py:5306] Model loading took 3.89 GiB memory and 2.709324 seconds
+(EngineCore pid=318295) INFO 07-15 19:57:40 [gpu_worker.py:538] Available KV cache memory: 15.83 GiB
+(EngineCore pid=318295) INFO 07-15 19:57:40 [kv_cache_utils.py:2146] GPU KV cache size: 129,648 tokens
+(EngineCore pid=318295) INFO 07-15 19:57:40 [kv_cache_utils.py:2147] Maximum concurrency for 2,048 tokens per request: 63.30x
+(EngineCore pid=318295) INFO 07-15 19:57:40 [cutedsl_warmup.py:97] Skipping CuTeDSL warmup because no compile units were requested.
+(EngineCore pid=318295) INFO 07-15 19:57:40 [jit_monitor.py:73] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
+(EngineCore pid=318295) INFO 07-15 19:57:40 [core.py:344] init engine (profile, create kv cache, warmup model) took 2.04 s
+(EngineCore pid=318295) INFO 07-15 19:57:41 [vllm.py:1042] Asynchronous scheduling is enabled.
+(EngineCore pid=318295) WARNING 07-15 19:57:41 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(EngineCore pid=318295) WARNING 07-15 19:57:41 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(EngineCore pid=318295) INFO 07-15 19:57:41 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(EngineCore pid=318295) INFO 07-15 19:57:41 [vllm.py:1322] Cudagraph is disabled under eager mode
+(EngineCore pid=318295) INFO 07-15 19:57:41 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(APIServer pid=318177) INFO 07-15 19:57:41 [api_server.py:612] Supported tasks: ['generate']
+(APIServer pid=318177) WARNING 07-15 19:57:41 [__init__.py:36] SECURITY WARNING: Development endpoints are enabled! This should NOT be used in production!
+(APIServer pid=318177) INFO 07-15 19:57:41 [hf.py:548] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
+(APIServer pid=318177) INFO 07-15 19:57:41 [api_server.py:616] Starting vLLM server on http://127.0.0.1:8378
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:37] Available routes are:
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /openapi.json, Methods: HEAD, GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /docs, Methods: HEAD, GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: HEAD, GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /redoc, Methods: HEAD, GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /load, Methods: GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /version, Methods: GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /health, Methods: GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /metrics, Methods: GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /tokenize, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /detokenize, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /v1/models, Methods: GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /ping, Methods: GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /ping, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /invocations, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /reset_prefix_cache, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /reset_mm_cache, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /reset_encoder_cache, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /pause, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /resume, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /is_paused, Methods: GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /init_weight_transfer_engine, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /start_weight_update, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /update_weights, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /finish_weight_update, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /get_world_size, Methods: GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /collective_rpc, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /server_info, Methods: GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /sleep, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /wake_up, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /is_sleeping, Methods: GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /v1/chat/completions/batch, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /v1/responses, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /v1/completions, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /v1/messages, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /generative_scoring, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /v1/completions/render, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /v1/chat/completions/derender, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /v1/completions/derender, Methods: POST
+(APIServer pid=318177) INFO 07-15 19:57:41 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
+(APIServer pid=318177) INFO:     Started server process [318177]
+(APIServer pid=318177) INFO:     Waiting for application startup.
+(APIServer pid=318177) INFO:     Application startup complete.
+(APIServer pid=318177) INFO:     127.0.0.1:39916 - "GET /health HTTP/1.1" 200 OK
+(EngineCore pid=318295) WARNING 07-15 19:57:43 [jit_monitor.py:129] Triton kernel JIT compilation during inference: _build_route_rows. This causes a latency spike; consider extending warmup to cover this shape/config.
+(APIServer pid=318177) INFO 07-15 19:57:51 [loggers.py:273] Engine 000: Avg prompt throughput: 2048.3 tokens/s, Avg generation throughput: 2241.9 tokens/s, Running: 256 reqs, Waiting: 1059 reqs, GPU KV cache usage: 36.1%, Prefix cache hit rate: 93.1%
+(APIServer pid=318177) INFO 07-15 19:58:01 [loggers.py:273] Engine 000: Avg prompt throughput: 659.1 tokens/s, Avg generation throughput: 2960.6 tokens/s, Running: 255 reqs, Waiting: 972 reqs, GPU KV cache usage: 47.9%, Prefix cache hit rate: 93.2%
+(APIServer pid=318177) INFO 07-15 19:58:11 [loggers.py:273] Engine 000: Avg prompt throughput: 1056.9 tokens/s, Avg generation throughput: 2930.4 tokens/s, Running: 255 reqs, Waiting: 841 reqs, GPU KV cache usage: 46.2%, Prefix cache hit rate: 93.2%
+(APIServer pid=318177) INFO 07-15 19:58:21 [loggers.py:273] Engine 000: Avg prompt throughput: 745.0 tokens/s, Avg generation throughput: 2933.8 tokens/s, Running: 256 reqs, Waiting: 742 reqs, GPU KV cache usage: 50.3%, Prefix cache hit rate: 93.3%
+(APIServer pid=318177) INFO 07-15 19:58:31 [loggers.py:273] Engine 000: Avg prompt throughput: 1024.4 tokens/s, Avg generation throughput: 2930.4 tokens/s, Running: 256 reqs, Waiting: 612 reqs, GPU KV cache usage: 44.9%, Prefix cache hit rate: 93.3%
+(APIServer pid=318177) INFO 07-15 19:58:41 [loggers.py:273] Engine 000: Avg prompt throughput: 879.9 tokens/s, Avg generation throughput: 2932.2 tokens/s, Running: 256 reqs, Waiting: 501 reqs, GPU KV cache usage: 46.7%, Prefix cache hit rate: 93.3%
+(APIServer pid=318177) INFO 07-15 19:58:51 [loggers.py:273] Engine 000: Avg prompt throughput: 1021.1 tokens/s, Avg generation throughput: 2930.9 tokens/s, Running: 256 reqs, Waiting: 372 reqs, GPU KV cache usage: 44.5%, Prefix cache hit rate: 93.4%
+(APIServer pid=318177) INFO 07-15 19:59:01 [loggers.py:273] Engine 000: Avg prompt throughput: 870.8 tokens/s, Avg generation throughput: 2958.9 tokens/s, Running: 255 reqs, Waiting: 268 reqs, GPU KV cache usage: 48.1%, Prefix cache hit rate: 93.4%
+(APIServer pid=318177) INFO 07-15 19:59:11 [loggers.py:273] Engine 000: Avg prompt throughput: 916.4 tokens/s, Avg generation throughput: 2931.6 tokens/s, Running: 255 reqs, Waiting: 147 reqs, GPU KV cache usage: 46.7%, Prefix cache hit rate: 93.4%
+(APIServer pid=318177) INFO 07-15 19:59:21 [loggers.py:273] Engine 000: Avg prompt throughput: 835.1 tokens/s, Avg generation throughput: 2934.0 tokens/s, Running: 255 reqs, Waiting: 50 reqs, GPU KV cache usage: 51.6%, Prefix cache hit rate: 93.4%
+(APIServer pid=318177) INFO 07-15 19:59:31 [loggers.py:273] Engine 000: Avg prompt throughput: 389.9 tokens/s, Avg generation throughput: 2849.7 tokens/s, Running: 191 reqs, Waiting: 0 reqs, GPU KV cache usage: 45.7%, Prefix cache hit rate: 93.4%
+(APIServer pid=318177) INFO 07-15 19:59:41 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1991.2 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 93.4%
+(APIServer pid=318177) INFO:     127.0.0.1:39930 - "POST /v1/completions HTTP/1.1" 200 OK
+(EngineCore pid=318295) INFO 07-15 19:59:47 [core.py:1214] [shutdown] EngineCore: trigger received signal=SIGTERM
+(APIServer pid=318177) INFO 07-15 19:59:47 [launcher.py:100] [shutdown] API server: shutdown triggered
+(APIServer pid=318177) INFO 07-15 19:59:47 [launcher.py:116] [shutdown] API server: stopping engine client mode=abort timeout=0s
+(EngineCore pid=318295) INFO 07-15 19:59:47 [core.py:1333] [shutdown] EngineCore: start mode=abort timeout=0s
+(EngineCore pid=318295) INFO 07-15 19:59:47 [core.py:1364] [shutdown] EngineCore: request processing complete; starting resource teardown
+(EngineCore pid=318295) INFO 07-15 19:59:47 [core.py:1227] [shutdown] EngineCore: exiting busy loop
+(APIServer pid=318177) INFO 07-15 19:59:47 [core_client.py:655] [shutdown] MPClient: start timeout=0s
+(APIServer pid=318177) INFO 07-15 19:59:47 [core_client.py:657] [shutdown] MPClient: stopping engine manager
+(APIServer pid=318177) WARNING 07-15 19:59:47 [utils.py:626] [shutdown] Process manager: force killing remaining processes count=1
+(APIServer pid=318177) INFO 07-15 19:59:47 [core_client.py:659] [shutdown] MPClient: engine manager stopped
+(APIServer pid=318177) INFO 07-15 19:59:47 [core_client.py:660] [shutdown] MPClient: cleaning up background resources
+(APIServer pid=318177) INFO 07-15 19:59:47 [core_client.py:662] [shutdown] MPClient: complete
+(APIServer pid=318177) INFO 07-15 19:59:47 [launcher.py:125] [shutdown] API server: engine client stopped
+(APIServer pid=318177) INFO 07-15 19:59:47 [launcher.py:128] [shutdown] API server: signalling HTTP server shutdown
+(APIServer pid=318177) INFO 07-15 19:59:47 [launcher.py:149] [shutdown] API server: shutting down FastAPI HTTP server
+(APIServer pid=318177) INFO:     Shutting down
+(APIServer pid=318177) INFO:     Waiting for application shutdown.
+(APIServer pid=318177) INFO:     Application shutdown complete.
+/home/henry/.local/share/uv/python/cpython-3.12.12-linux-x86_64-gnu/lib/python3.12/multiprocessing/resource_tracker.py:279: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
+  warnings.warn('resource_tracker: There appear to be %d '

evals/healing_breadth/keep25_traj/step0500_chat.json.server.log

@@ -0,0 +1,277 @@
+(APIServer pid=319610) INFO 07-15 20:00:12 [api_utils.py:339] 
+(APIServer pid=319610) INFO 07-15 20:00:12 [api_utils.py:339]        █     █     █▄   ▄█
+(APIServer pid=319610) INFO 07-15 20:00:12 [api_utils.py:339]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.25.0
+(APIServer pid=319610) INFO 07-15 20:00:12 [api_utils.py:339]   █▄█▀ █     █     █     █  model   outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0500
+(APIServer pid=319610) INFO 07-15 20:00:12 [api_utils.py:339]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
+(APIServer pid=319610) INFO 07-15 20:00:12 [api_utils.py:339] 
+(APIServer pid=319610) INFO 07-15 20:00:12 [api_utils.py:273] non-default args: {'model_tag': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0500', 'host': '127.0.0.1', 'port': 8378, 'model': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0500', 'max_model_len': 2048, 'enforce_eager': True, 'served_model_name': ['student'], 'gpu_memory_utilization': 0.85}
+(APIServer pid=319610) INFO 07-15 20:00:12 [model.py:619] Resolved architecture: PrunedOlmoeForCausalLM
+(APIServer pid=319610) INFO 07-15 20:00:12 [model.py:1776] Using max model len 2048
+(APIServer pid=319610) INFO 07-15 20:00:12 [vllm.py:1042] Asynchronous scheduling is enabled.
+(APIServer pid=319610) WARNING 07-15 20:00:12 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(APIServer pid=319610) WARNING 07-15 20:00:12 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(APIServer pid=319610) INFO 07-15 20:00:12 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(APIServer pid=319610) INFO 07-15 20:00:12 [vllm.py:1322] Cudagraph is disabled under eager mode
+(APIServer pid=319610) INFO 07-15 20:00:12 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(EngineCore pid=319728) INFO 07-15 20:00:19 [core.py:114] Initializing a V1 LLM engine (v0.25.0) with config: model='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0500', speculative_config=None, tokenizer='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0500', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=student, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.NONE: 0>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, moe_backend='auto', linear_backend='auto')
+(EngineCore pid=319728) INFO 07-15 20:00:19 [parallel_state.py:1607] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.0.15:54145 backend=nccl
+(EngineCore pid=319728) INFO 07-15 20:00:19 [parallel_state.py:1942] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A
+(EngineCore pid=319728) INFO 07-15 20:00:20 [topk_topp_sampler.py:55] Using FlashInfer for top-p & top-k sampling.
+(EngineCore pid=319728) INFO 07-15 20:00:20 [gpu_model_runner.py:5209] Starting to load model outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0500...
+(EngineCore pid=319728) INFO 07-15 20:00:21 [cuda.py:476] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
+(EngineCore pid=319728) INFO 07-15 20:00:21 [flash_attn.py:718] Using FlashAttention version 2
+(EngineCore pid=319728) /home/henry/.cache/glean/megablocks-variable-93a1479bc15b/megablocks/grouped_gemm_util.py:10: UserWarning: Grouped GEMM not available.
+(EngineCore pid=319728)   warnings.warn('Grouped GEMM not available.')
+(EngineCore pid=319728) INFO 07-15 20:00:21 [weight_utils.py:849] Filesystem type for checkpoints: EXT4. Checkpoint size: 3.89 GiB. Available RAM: 106.42 GiB.
+(EngineCore pid=319728) INFO 07-15 20:00:21 [weight_utils.py:872] Auto-prefetch is disabled because the filesystem (EXT4) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.
+(EngineCore pid=319728) 
+(EngineCore pid=319728) 
+(EngineCore pid=319728) 
+(EngineCore pid=319728) 
+(EngineCore pid=319728) INFO 07-15 20:00:23 [default_loader.py:430] Loading weights took 2.57 seconds
+(EngineCore pid=319728) INFO 07-15 20:00:24 [gpu_model_runner.py:5306] Model loading took 3.89 GiB memory and 2.751591 seconds
+(EngineCore pid=319728) INFO 07-15 20:00:25 [gpu_worker.py:538] Available KV cache memory: 15.82 GiB
+(EngineCore pid=319728) INFO 07-15 20:00:25 [kv_cache_utils.py:2146] GPU KV cache size: 129,584 tokens
+(EngineCore pid=319728) INFO 07-15 20:00:25 [kv_cache_utils.py:2147] Maximum concurrency for 2,048 tokens per request: 63.27x
+(EngineCore pid=319728) INFO 07-15 20:00:26 [cutedsl_warmup.py:97] Skipping CuTeDSL warmup because no compile units were requested.
+(EngineCore pid=319728) INFO 07-15 20:00:26 [jit_monitor.py:73] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
+(EngineCore pid=319728) INFO 07-15 20:00:26 [core.py:344] init engine (profile, create kv cache, warmup model) took 2.03 s
+(EngineCore pid=319728) INFO 07-15 20:00:26 [vllm.py:1042] Asynchronous scheduling is enabled.
+(EngineCore pid=319728) WARNING 07-15 20:00:26 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(EngineCore pid=319728) WARNING 07-15 20:00:26 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(EngineCore pid=319728) INFO 07-15 20:00:26 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(EngineCore pid=319728) INFO 07-15 20:00:26 [vllm.py:1322] Cudagraph is disabled under eager mode
+(EngineCore pid=319728) INFO 07-15 20:00:26 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(APIServer pid=319610) INFO 07-15 20:00:26 [api_server.py:612] Supported tasks: ['generate']
+(APIServer pid=319610) WARNING 07-15 20:00:26 [__init__.py:36] SECURITY WARNING: Development endpoints are enabled! This should NOT be used in production!
+(APIServer pid=319610) INFO 07-15 20:00:26 [hf.py:548] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
+(APIServer pid=319610) INFO 07-15 20:00:26 [api_server.py:616] Starting vLLM server on http://127.0.0.1:8378
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:37] Available routes are:
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /openapi.json, Methods: HEAD, GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /docs, Methods: HEAD, GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: HEAD, GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /redoc, Methods: HEAD, GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /load, Methods: GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /version, Methods: GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /health, Methods: GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /metrics, Methods: GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /tokenize, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /detokenize, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /v1/models, Methods: GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /ping, Methods: GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /ping, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /invocations, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /reset_prefix_cache, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /reset_mm_cache, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /reset_encoder_cache, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /pause, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /resume, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /is_paused, Methods: GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /init_weight_transfer_engine, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /start_weight_update, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /update_weights, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /finish_weight_update, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /get_world_size, Methods: GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /collective_rpc, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /server_info, Methods: GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /sleep, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /wake_up, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /is_sleeping, Methods: GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /v1/chat/completions/batch, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /v1/responses, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /v1/completions, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /v1/messages, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /generative_scoring, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /v1/completions/render, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /v1/chat/completions/derender, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /v1/completions/derender, Methods: POST
+(APIServer pid=319610) INFO 07-15 20:00:26 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
+(APIServer pid=319610) INFO:     Started server process [319610]
+(APIServer pid=319610) INFO:     Waiting for application startup.
+(APIServer pid=319610) INFO:     Application startup complete.
+(APIServer pid=319610) INFO:     127.0.0.1:52108 - "GET /health HTTP/1.1" 200 OK
+(EngineCore pid=319728) WARNING 07-15 20:00:28 [jit_monitor.py:129] Triton kernel JIT compilation during inference: _build_route_rows. This causes a latency spike; consider extending warmup to cover this shape/config.
+(APIServer pid=319610) INFO 07-15 20:00:37 [loggers.py:273] Engine 000: Avg prompt throughput: 2023.7 tokens/s, Avg generation throughput: 2223.4 tokens/s, Running: 256 reqs, Waiting: 1062 reqs, GPU KV cache usage: 36.0%, Prefix cache hit rate: 93.1%
+(APIServer pid=319610) INFO 07-15 20:00:47 [loggers.py:273] Engine 000: Avg prompt throughput: 686.0 tokens/s, Avg generation throughput: 2960.6 tokens/s, Running: 255 reqs, Waiting: 973 reqs, GPU KV cache usage: 47.6%, Prefix cache hit rate: 93.2%
+(APIServer pid=319610) INFO 07-15 20:00:57 [loggers.py:273] Engine 000: Avg prompt throughput: 948.2 tokens/s, Avg generation throughput: 2906.7 tokens/s, Running: 255 reqs, Waiting: 856 reqs, GPU KV cache usage: 48.4%, Prefix cache hit rate: 93.2%
+(APIServer pid=319610) INFO 07-15 20:01:07 [loggers.py:273] Engine 000: Avg prompt throughput: 713.6 tokens/s, Avg generation throughput: 2934.3 tokens/s, Running: 255 reqs, Waiting: 760 reqs, GPU KV cache usage: 50.9%, Prefix cache hit rate: 93.3%
+(APIServer pid=319610) INFO 07-15 20:01:17 [loggers.py:273] Engine 000: Avg prompt throughput: 1006.1 tokens/s, Avg generation throughput: 2904.1 tokens/s, Running: 256 reqs, Waiting: 630 reqs, GPU KV cache usage: 44.5%, Prefix cache hit rate: 93.3%
+(APIServer pid=319610) INFO 07-15 20:01:27 [loggers.py:273] Engine 000: Avg prompt throughput: 868.7 tokens/s, Avg generation throughput: 2907.5 tokens/s, Running: 255 reqs, Waiting: 521 reqs, GPU KV cache usage: 46.3%, Prefix cache hit rate: 93.3%
+(APIServer pid=319610) INFO 07-15 20:01:37 [loggers.py:273] Engine 000: Avg prompt throughput: 875.0 tokens/s, Avg generation throughput: 2957.9 tokens/s, Running: 256 reqs, Waiting: 410 reqs, GPU KV cache usage: 46.8%, Prefix cache hit rate: 93.4%
+(APIServer pid=319610) INFO 07-15 20:01:47 [loggers.py:273] Engine 000: Avg prompt throughput: 931.2 tokens/s, Avg generation throughput: 2907.0 tokens/s, Running: 255 reqs, Waiting: 300 reqs, GPU KV cache usage: 48.5%, Prefix cache hit rate: 93.4%
+(APIServer pid=319610) INFO 07-15 20:01:57 [loggers.py:273] Engine 000: Avg prompt throughput: 914.1 tokens/s, Avg generation throughput: 2906.6 tokens/s, Running: 256 reqs, Waiting: 184 reqs, GPU KV cache usage: 47.3%, Prefix cache hit rate: 93.4%
+(APIServer pid=319610) INFO 07-15 20:02:07 [loggers.py:273] Engine 000: Avg prompt throughput: 844.5 tokens/s, Avg generation throughput: 2933.1 tokens/s, Running: 254 reqs, Waiting: 77 reqs, GPU KV cache usage: 48.0%, Prefix cache hit rate: 93.4%
+(APIServer pid=320368) INFO 07-15 20:02:10 [api_utils.py:339] 
+(APIServer pid=320368) INFO 07-15 20:02:10 [api_utils.py:339]        █     █     █▄   ▄█
+(APIServer pid=320368) INFO 07-15 20:02:10 [api_utils.py:339]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.25.0
+(APIServer pid=320368) INFO 07-15 20:02:10 [api_utils.py:339]   █▄█▀ █     █     █     █  model   outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0500
+(APIServer pid=320368) INFO 07-15 20:02:10 [api_utils.py:339]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
+(APIServer pid=320368) INFO 07-15 20:02:10 [api_utils.py:339] 
+(APIServer pid=320368) INFO 07-15 20:02:10 [api_utils.py:273] non-default args: {'model_tag': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0500', 'host': '127.0.0.1', 'port': 8377, 'model': 'outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0500', 'max_model_len': 2048, 'enforce_eager': True, 'served_model_name': ['student'], 'gpu_memory_utilization': 0.85}
+(APIServer pid=320368) INFO 07-15 20:02:10 [model.py:619] Resolved architecture: PrunedOlmoeForCausalLM
+(APIServer pid=320368) INFO 07-15 20:02:10 [model.py:1776] Using max model len 2048
+(APIServer pid=320368) INFO 07-15 20:02:10 [vllm.py:1042] Asynchronous scheduling is enabled.
+(APIServer pid=320368) WARNING 07-15 20:02:10 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(APIServer pid=320368) WARNING 07-15 20:02:10 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(APIServer pid=320368) INFO 07-15 20:02:10 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(APIServer pid=320368) INFO 07-15 20:02:10 [vllm.py:1322] Cudagraph is disabled under eager mode
+(APIServer pid=320368) INFO 07-15 20:02:10 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(APIServer pid=319610) INFO 07-15 20:02:17 [loggers.py:273] Engine 000: Avg prompt throughput: 638.4 tokens/s, Avg generation throughput: 2932.7 tokens/s, Running: 248 reqs, Waiting: 0 reqs, GPU KV cache usage: 52.4%, Prefix cache hit rate: 93.4%
+(EngineCore pid=320486) INFO 07-15 20:02:17 [core.py:114] Initializing a V1 LLM engine (v0.25.0) with config: model='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0500', speculative_config=None, tokenizer='outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0500', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=student, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.NONE: 0>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, moe_backend='auto', linear_backend='auto')
+(EngineCore pid=320486) INFO 07-15 20:02:18 [parallel_state.py:1607] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.0.15:40153 backend=nccl
+(EngineCore pid=320486) INFO 07-15 20:02:18 [parallel_state.py:1942] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A
+(EngineCore pid=320486) INFO 07-15 20:02:19 [topk_topp_sampler.py:55] Using FlashInfer for top-p & top-k sampling.
+(EngineCore pid=320486) INFO 07-15 20:02:19 [gpu_model_runner.py:5209] Starting to load model outputs/healed/healing_breadth/glean_math_keep25_seed1224_long768/step0500...
+(EngineCore pid=320486) INFO 07-15 20:02:19 [cuda.py:476] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
+(EngineCore pid=320486) INFO 07-15 20:02:19 [flash_attn.py:718] Using FlashAttention version 2
+(EngineCore pid=320486) /home/henry/.cache/glean/megablocks-variable-93a1479bc15b/megablocks/grouped_gemm_util.py:10: UserWarning: Grouped GEMM not available.
+(EngineCore pid=320486)   warnings.warn('Grouped GEMM not available.')
+(EngineCore pid=320486) INFO 07-15 20:02:19 [weight_utils.py:849] Filesystem type for checkpoints: EXT4. Checkpoint size: 3.89 GiB. Available RAM: 106.39 GiB.
+(EngineCore pid=320486) INFO 07-15 20:02:19 [weight_utils.py:872] Auto-prefetch is disabled because the filesystem (EXT4) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.
+(EngineCore pid=320486) 
+(EngineCore pid=320486) 
+(EngineCore pid=320486) 
+(EngineCore pid=320486) 
+(EngineCore pid=320486) INFO 07-15 20:02:22 [default_loader.py:430] Loading weights took 2.54 seconds
+(EngineCore pid=320486) INFO 07-15 20:02:22 [gpu_model_runner.py:5306] Model loading took 3.89 GiB memory and 2.722867 seconds
+(EngineCore pid=320486) INFO 07-15 20:02:24 [gpu_worker.py:538] Available KV cache memory: 15.82 GiB
+(EngineCore pid=320486) INFO 07-15 20:02:24 [kv_cache_utils.py:2146] GPU KV cache size: 129,584 tokens
+(EngineCore pid=320486) INFO 07-15 20:02:24 [kv_cache_utils.py:2147] Maximum concurrency for 2,048 tokens per request: 63.27x
+(EngineCore pid=320486) INFO 07-15 20:02:24 [cutedsl_warmup.py:97] Skipping CuTeDSL warmup because no compile units were requested.
+(EngineCore pid=320486) INFO 07-15 20:02:24 [jit_monitor.py:73] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
+(EngineCore pid=320486) INFO 07-15 20:02:24 [core.py:344] init engine (profile, create kv cache, warmup model) took 2.14 s
+(EngineCore pid=320486) INFO 07-15 20:02:24 [vllm.py:1042] Asynchronous scheduling is enabled.
+(EngineCore pid=320486) WARNING 07-15 20:02:24 [vllm.py:1096] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(EngineCore pid=320486) WARNING 07-15 20:02:24 [vllm.py:1144] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(EngineCore pid=320486) INFO 07-15 20:02:24 [kernel.py:292] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(EngineCore pid=320486) INFO 07-15 20:02:24 [vllm.py:1322] Cudagraph is disabled under eager mode
+(EngineCore pid=320486) INFO 07-15 20:02:24 [compilation.py:312] Enabled custom fusions: norm_quant, act_quant
+(APIServer pid=320368) INFO 07-15 20:02:24 [api_server.py:612] Supported tasks: ['generate']
+(APIServer pid=320368) WARNING 07-15 20:02:24 [__init__.py:36] SECURITY WARNING: Development endpoints are enabled! This should NOT be used in production!
+(APIServer pid=320368) INFO 07-15 20:02:25 [hf.py:548] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
+(APIServer pid=320368) INFO 07-15 20:02:25 [api_server.py:616] Starting vLLM server on http://127.0.0.1:8377
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:37] Available routes are:
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /openapi.json, Methods: HEAD, GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /docs, Methods: HEAD, GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: HEAD, GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /redoc, Methods: HEAD, GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /load, Methods: GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /version, Methods: GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /health, Methods: GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /metrics, Methods: GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /tokenize, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /detokenize, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /v1/models, Methods: GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /ping, Methods: GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /ping, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /invocations, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /reset_prefix_cache, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /reset_mm_cache, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /reset_encoder_cache, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /pause, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /resume, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /is_paused, Methods: GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /init_weight_transfer_engine, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /start_weight_update, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /update_weights, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /finish_weight_update, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /get_world_size, Methods: GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /collective_rpc, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /server_info, Methods: GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /sleep, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /wake_up, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /is_sleeping, Methods: GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /v1/chat/completions/batch, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /v1/responses, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /v1/completions, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /v1/messages, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /generative_scoring, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /v1/completions/render, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /v1/chat/completions/derender, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /v1/completions/derender, Methods: POST
+(APIServer pid=320368) INFO 07-15 20:02:25 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
+(APIServer pid=320368) INFO:     Started server process [320368]
+(APIServer pid=320368) INFO:     Waiting for application startup.
+(APIServer pid=320368) INFO:     Application startup complete.
+(APIServer pid=320368) INFO:     127.0.0.1:37372 - "GET /health HTTP/1.1" 200 OK
+(EngineCore pid=320486) WARNING 07-15 20:02:26 [jit_monitor.py:129] Triton kernel JIT compilation during inference: _build_route_rows. This causes a latency spike; consider extending warmup to cover this shape/config.
+(APIServer pid=319610) INFO 07-15 20:02:27 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2556.8 tokens/s, Running: 93 reqs, Waiting: 0 reqs, GPU KV cache usage: 27.5%, Prefix cache hit rate: 93.4%
+(APIServer pid=319610) INFO:     127.0.0.1:52118 - "POST /v1/completions HTTP/1.1" 200 OK
+(APIServer pid=320368) INFO 07-15 20:02:35 [loggers.py:273] Engine 000: Avg prompt throughput: 2028.0 tokens/s, Avg generation throughput: 2313.6 tokens/s, Running: 255 reqs, Waiting: 1062 reqs, GPU KV cache usage: 36.6%, Prefix cache hit rate: 93.1%
+(EngineCore pid=319728) INFO 07-15 20:02:35 [core.py:1214] [shutdown] EngineCore: trigger received signal=SIGTERM
+(APIServer pid=319610) INFO 07-15 20:02:35 [launcher.py:100] [shutdown] API server: shutdown triggered
+(APIServer pid=319610) INFO 07-15 20:02:35 [launcher.py:116] [shutdown] API server: stopping engine client mode=abort timeout=0s
+(EngineCore pid=319728) INFO 07-15 20:02:35 [core.py:1333] [shutdown] EngineCore: start mode=abort timeout=0s
+(EngineCore pid=319728) INFO 07-15 20:02:35 [core.py:1364] [shutdown] EngineCore: request processing complete; starting resource teardown
+(EngineCore pid=319728) INFO 07-15 20:02:35 [core.py:1227] [shutdown] EngineCore: exiting busy loop
+(APIServer pid=319610) INFO 07-15 20:02:35 [core_client.py:655] [shutdown] MPClient: start timeout=0s
+(APIServer pid=319610) INFO 07-15 20:02:35 [core_client.py:657] [shutdown] MPClient: stopping engine manager
+(APIServer pid=319610) WARNING 07-15 20:02:35 [utils.py:626] [shutdown] Process manager: force killing remaining processes count=1
+(APIServer pid=319610) INFO:     Shutting down
+(APIServer pid=319610) INFO 07-15 20:02:35 [core_client.py:659] [shutdown] MPClient: engine manager stopped
+(APIServer pid=319610) INFO 07-15 20:02:35 [core_client.py:660] [shutdown] MPClient: cleaning up background resources
+(APIServer pid=319610) INFO 07-15 20:02:35 [core_client.py:662] [shutdown] MPClient: complete
+(APIServer pid=319610) INFO 07-15 20:02:35 [launcher.py:125] [shutdown] API server: engine client stopped
+(APIServer pid=319610) INFO 07-15 20:02:35 [launcher.py:128] [shutdown] API server: signalling HTTP server shutdown
+(APIServer pid=319610) INFO 07-15 20:02:35 [launcher.py:149] [shutdown] API server: shutting down FastAPI HTTP server
+(APIServer pid=319610) INFO:     Shutting down
+(APIServer pid=319610) INFO:     Waiting for application shutdown.
+(APIServer pid=319610) INFO:     Application shutdown complete.
+/home/henry/.local/share/uv/python/cpython-3.12.12-linux-x86_64-gnu/lib/python3.12/multiprocessing/resource_tracker.py:279: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
+  warnings.warn('resource_tracker: There appear to be %d '
+(APIServer pid=320368) INFO 07-15 20:02:45 [loggers.py:273] Engine 000: Avg prompt throughput: 671.6 tokens/s, Avg generation throughput: 2960.1 tokens/s, Running: 255 reqs, Waiting: 974 reqs, GPU KV cache usage: 48.3%, Prefix cache hit rate: 93.2%
+(APIServer pid=320368) INFO 07-15 20:02:55 [loggers.py:273] Engine 000: Avg prompt throughput: 999.0 tokens/s, Avg generation throughput: 2904.8 tokens/s, Running: 256 reqs, Waiting: 849 reqs, GPU KV cache usage: 47.2%, Prefix cache hit rate: 93.2%
+(APIServer pid=320368) INFO 07-15 20:03:05 [loggers.py:273] Engine 000: Avg prompt throughput: 697.6 tokens/s, Avg generation throughput: 2934.7 tokens/s, Running: 255 reqs, Waiting: 757 reqs, GPU KV cache usage: 50.6%, Prefix cache hit rate: 93.3%
+(APIServer pid=320368) INFO 07-15 20:03:15 [loggers.py:273] Engine 000: Avg prompt throughput: 993.8 tokens/s, Avg generation throughput: 2879.3 tokens/s, Running: 254 reqs, Waiting: 628 reqs, GPU KV cache usage: 43.7%, Prefix cache hit rate: 93.3%
+(APIServer pid=320368) INFO 07-15 20:03:25 [loggers.py:273] Engine 000: Avg prompt throughput: 894.6 tokens/s, Avg generation throughput: 2932.7 tokens/s, Running: 256 reqs, Waiting: 517 reqs, GPU KV cache usage: 46.5%, Prefix cache hit rate: 93.3%
+(APIServer pid=320368) INFO 07-15 20:03:35 [loggers.py:273] Engine 000: Avg prompt throughput: 844.9 tokens/s, Avg generation throughput: 2958.0 tokens/s, Running: 256 reqs, Waiting: 406 reqs, GPU KV cache usage: 47.1%, Prefix cache hit rate: 93.4%
+(APIServer pid=320368) INFO 07-15 20:03:45 [loggers.py:273] Engine 000: Avg prompt throughput: 986.5 tokens/s, Avg generation throughput: 2931.6 tokens/s, Running: 256 reqs, Waiting: 289 reqs, GPU KV cache usage: 46.6%, Prefix cache hit rate: 93.4%
+(APIServer pid=320368) INFO 07-15 20:03:55 [loggers.py:273] Engine 000: Avg prompt throughput: 886.9 tokens/s, Avg generation throughput: 2932.8 tokens/s, Running: 255 reqs, Waiting: 178 reqs, GPU KV cache usage: 47.7%, Prefix cache hit rate: 93.4%
+(APIServer pid=320368) INFO 07-15 20:04:05 [loggers.py:273] Engine 000: Avg prompt throughput: 810.9 tokens/s, Avg generation throughput: 2907.8 tokens/s, Running: 255 reqs, Waiting: 75 reqs, GPU KV cache usage: 49.9%, Prefix cache hit rate: 93.4%
+(APIServer pid=320368) INFO 07-15 20:04:15 [loggers.py:273] Engine 000: Avg prompt throughput: 632.9 tokens/s, Avg generation throughput: 2922.2 tokens/s, Running: 237 reqs, Waiting: 0 reqs, GPU KV cache usage: 50.8%, Prefix cache hit rate: 93.4%
+(APIServer pid=320368) INFO 07-15 20:04:25 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2517.4 tokens/s, Running: 76 reqs, Waiting: 0 reqs, GPU KV cache usage: 23.7%, Prefix cache hit rate: 93.4%
+(APIServer pid=320368) INFO:     127.0.0.1:37374 - "POST /v1/completions HTTP/1.1" 200 OK
+(APIServer pid=320368) INFO 07-15 20:04:35 [loggers.py:273] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 745.7 tokens/s, Running: 0 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 93.4%
+(EngineCore pid=320486) INFO 07-15 20:04:35 [core.py:1214] [shutdown] EngineCore: trigger received signal=SIGTERM
+(APIServer pid=320368) INFO 07-15 20:04:35 [launcher.py:100] [shutdown] API server: shutdown triggered
+(APIServer pid=320368) INFO 07-15 20:04:35 [launcher.py:116] [shutdown] API server: stopping engine client mode=abort timeout=0s
+(EngineCore pid=320486) INFO 07-15 20:04:35 [core.py:1333] [shutdown] EngineCore: start mode=abort timeout=0s
+(EngineCore pid=320486) INFO 07-15 20:04:35 [core.py:1364] [shutdown] EngineCore: request processing complete; starting resource teardown
+(EngineCore pid=320486) INFO 07-15 20:04:35 [core.py:1227] [shutdown] EngineCore: exiting busy loop
+(APIServer pid=320368) INFO 07-15 20:04:35 [core_client.py:655] [shutdown] MPClient: start timeout=0s
+(APIServer pid=320368) INFO 07-15 20:04:35 [core_client.py:657] [shutdown] MPClient: stopping engine manager
+(APIServer pid=320368) WARNING 07-15 20:04:35 [utils.py:626] [shutdown] Process manager: force killing remaining processes count=1
+(APIServer pid=320368) INFO:     Shutting down
+(APIServer pid=320368) INFO 07-15 20:04:35 [core_client.py:659] [shutdown] MPClient: engine manager stopped
+(APIServer pid=320368) INFO 07-15 20:04:35 [core_client.py:660] [shutdown] MPClient: cleaning up background resources
+(APIServer pid=320368) INFO 07-15 20:04:35 [core_client.py:662] [shutdown] MPClient: complete
+(APIServer pid=320368) ERROR 07-15 20:04:35 [async_llm.py:704] AsyncLLM output_handler failed.
+(APIServer pid=320368) ERROR 07-15 20:04:35 [async_llm.py:704] Traceback (most recent call last):
+(APIServer pid=320368) ERROR 07-15 20:04:35 [async_llm.py:704]   File "/home/henry/Documents/PythonProjects/variable-reap/vllm-plugin/.venv25/lib/python3.12/site-packages/vllm/v1/engine/async_llm.py", line 660, in output_handler
+(APIServer pid=320368) ERROR 07-15 20:04:35 [async_llm.py:704]     outputs = await engine_core.get_output_async()
+(APIServer pid=320368) ERROR 07-15 20:04:35 [async_llm.py:704]               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(APIServer pid=320368) ERROR 07-15 20:04:35 [async_llm.py:704]   File "/home/henry/Documents/PythonProjects/variable-reap/vllm-plugin/.venv25/lib/python3.12/site-packages/vllm/v1/engine/core_client.py", line 1061, in get_output_async
+(APIServer pid=320368) ERROR 07-15 20:04:35 [async_llm.py:704]     raise self._format_exception(outputs) from None
+(APIServer pid=320368) ERROR 07-15 20:04:35 [async_llm.py:704] vllm.v1.engine.exceptions.EngineDeadError: EngineCore encountered an issue. See stack trace (above) for the root cause.
+(APIServer pid=320368) INFO 07-15 20:04:35 [launcher.py:125] [shutdown] API server: engine client stopped
+(APIServer pid=320368) INFO 07-15 20:04:35 [launcher.py:128] [shutdown] API server: signalling HTTP server shutdown
+(APIServer pid=320368) INFO 07-15 20:04:35 [launcher.py:149] [shutdown] API server: shutting down FastAPI HTTP server
+(APIServer pid=320368) INFO:     Shutting down
+(APIServer pid=320368) INFO:     Waiting for application shutdown.
+(APIServer pid=320368) INFO:     Application shutdown complete.
+/home/henry/.local/share/uv/python/cpython-3.12.12-linux-x86_64-gnu/lib/python3.12/multiprocessing/resource_tracker.py:279: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
+  warnings.warn('resource_tracker: There appear to be %d '

evals/inst_baseline/allenai__OLMoE-1B-7B-0125-Instruct/results_2026-07-10T14-26-25.801129.json

@@ -0,0 +1,203 @@
+{
+  "results": {
+    "gsm8k_cot": {
+      "name": "gsm8k_cot",
+      "alias": "gsm8k_cot",
+      "sample_len": 250,
+      "exact_match,strict-match": 0.716,
+      "exact_match_stderr,strict-match": 0.028576958730437394,
+      "exact_match,flexible-extract": 0.716,
+      "exact_match_stderr,flexible-extract": 0.028576958730437394
+    }
+  },
+  "group_subtasks": {},
+  "configs": {
+    "gsm8k_cot": {
+      "task": "gsm8k_cot",
+      "dataset_path": "openai/gsm8k",
+      "dataset_name": "main",
+      "test_split": "test",
+      "doc_to_text": "Q: {{question}}\nA:",
+      "doc_to_target": "{{answer.split('####')[-1].strip() if answer is defined else target}}",
+      "unsafe_code": false,
+      "description": "",
+      "target_delimiter": " ",
+      "fewshot_delimiter": "\n\n",
+      "fewshot_config": {
+        "sampler": "first_n",
+        "split": null,
+        "process_docs": null,
+        "fewshot_indices": null,
+        "samples": [
+          {
+            "question": "There are 15 trees in the grove. Grove workers will plant trees in the grove today. After they are done, there will be 21 trees. How many trees did the grove workers plant today?",
+            "target": "There are 15 trees originally. Then there were 21 trees after some more were planted. So there must have been 21 - 15 = 6. The answer is 6."
+          },
+          {
+            "question": "If there are 3 cars in the parking lot and 2 more cars arrive, how many cars are in the parking lot?",
+            "target": "There are originally 3 cars. 2 more cars arrive. 3 + 2 = 5. The answer is 5."
+          },
+          {
+            "question": "Leah had 32 chocolates and her sister had 42. If they ate 35, how many pieces do they have left in total?",
+            "target": "Originally, Leah had 32 chocolates. Her sister had 42. So in total they had 32 + 42 = 74. After eating 35, they had 74 - 35 = 39. The answer is 39."
+          },
+          {
+            "question": "Jason had 20 lollipops. He gave Denny some lollipops. Now Jason has 12 lollipops. How many lollipops did Jason give to Denny?",
+            "target": "Jason started with 20 lollipops. Then he had 12 after giving some to Denny. So he gave Denny 20 - 12 = 8. The answer is 8."
+          },
+          {
+            "question": "Shawn has five toys. For Christmas, he got two toys each from his mom and dad. How many toys does he have now?",
+            "target": "Shawn started with 5 toys. If he got 2 toys each from his mom and dad, then that is 4 more toys. 5 + 4 = 9. The answer is 9."
+          },
+          {
+            "question": "There were nine computers in the server room. Five more computers were installed each day, from monday to thursday. How many computers are now in the server room?",
+            "target": "There were originally 9 computers. For each of 4 days, 5 more computers were added. So 5 * 4 = 20 computers were added. 9 + 20 is 29. The answer is 29."
+          },
+          {
+            "question": "Michael had 58 golf balls. On tuesday, he lost 23 golf balls. On wednesday, he lost 2 more. How many golf balls did he have at the end of wednesday?",
+            "target": "Michael started with 58 golf balls. After losing 23 on tuesday, he had 58 - 23 = 35. After losing 2 more, he had 35 - 2 = 33 golf balls. The answer is 33."
+          },
+          {
+            "question": "Olivia has $23. She bought five bagels for $3 each. How much money does she have left?",
+            "target": "Olivia had 23 dollars. 5 bagels for 3 dollars each will be 5 x 3 = 15 dollars. So she has 23 - 15 dollars left. 23 - 15 is 8. The answer is 8."
+          }
+        ],
+        "doc_to_text": "Q: {{question}}\nA:",
+        "doc_to_choice": null,
+        "doc_to_target": "{{answer.split('####')[-1].strip() if answer is defined else target}}",
+        "gen_prefix": null,
+        "fewshot_delimiter": "\n\n",
+        "target_delimiter": " "
+      },
+      "num_fewshot": 8,
+      "metric_list": [
+        {
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "ignore_case": true,
+          "ignore_punctuation": false,
+          "metric": "exact_match",
+          "regexes_to_ignore": [
+            ",",
+            "\\$",
+            "(?s).*#### ",
+            "\\.$"
+          ]
+        }
+      ],
+      "output_type": "generate_until",
+      "generation_kwargs": {
+        "do_sample": false,
+        "until": [
+          "Q:",
+          "</s>",
+          "<|im_end|>"
+        ]
+      },
+      "repeats": 1,
+      "filter_list": [
+        {
+          "filter": [
+            {
+              "function": "regex",
+              "regex_pattern": "The answer is (\\-?[0-9\\.\\,]+)."
+            },
+            {
+              "function": "take_first"
+            }
+          ],
+          "name": "strict-match"
+        },
+        {
+          "filter": [
+            {
+              "function": "regex",
+              "group_select": -1,
+              "regex_pattern": "(-?[$0-9.,]{2,})|(-?[0-9]+)"
+            },
+            {
+              "function": "take_first"
+            }
+          ],
+          "name": "flexible-extract"
+        }
+      ],
+      "should_decontaminate": false,
+      "metadata": {
+        "version": 3.0,
+        "pretrained": "allenai/OLMoE-1B-7B-0125-Instruct",
+        "dtype": "bfloat16",
+        "config_source": "/home/henry/Documents/PythonProjects/variable-reap/.venv/lib/python3.12/site-packages/lm_eval/tasks/gsm8k/gsm8k-cot.yaml"
+      }
+    }
+  },
+  "versions": {
+    "gsm8k_cot": 3.0
+  },
+  "n-shot": {
+    "gsm8k_cot": 8
+  },
+  "higher_is_better": {
+    "gsm8k_cot": {
+      "exact_match": true
+    }
+  },
+  "n-samples": {
+    "gsm8k_cot": {
+      "original": 1319,
+      "effective": 250
+    }
+  },
+  "config": {
+    "model": "hf",
+    "model_args": {
+      "pretrained": "allenai/OLMoE-1B-7B-0125-Instruct",
+      "dtype": "bfloat16"
+    },
+    "model_num_parameters": 6919161856,
+    "model_dtype": "torch.bfloat16",
+    "model_revision": "main",
+    "model_sha": "b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e",
+    "batch_size": "auto",
+    "batch_sizes": [],
+    "device": "cuda:0",
+    "use_cache": null,
+    "limit": 250.0,
+    "bootstrap_iters": 100000,
+    "gen_kwargs": {},
+    "random_seed": 0,
+    "numpy_seed": 1234,
+    "torch_seed": 1234,
+    "fewshot_seed": 1234
+  },
+  "git_hash": "42ce96a",
+  "date": 1783717435.720102,
+  "pretty_env_info": "PyTorch version: 2.12.1+cu130\nIs debug build: False\nCUDA used to build PyTorch: 13.0\nROCM used to build PyTorch: N/A\n\nOS: Pop!_OS 22.04 LTS (x86_64)\nGCC version: (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0\nClang version: Could not collect\nCMake version: version 3.22.1\nLibc version: glibc-2.35\n\nPython version: 3.12.12 (main, Oct 31 2025, 23:02:31) [Clang 21.1.4 ] (64-bit runtime)\nPython platform: Linux-7.0.11-76070011-generic-x86_64-with-glibc2.35\nIs CUDA available: True\nCUDA runtime version: 12.8.93\nCUDA_MODULE_LOADING set to: \nGPU models and configuration: \nGPU 0: NVIDIA GeForce RTX 3090\nGPU 1: NVIDIA GeForce RTX 3090\nGPU 2: NVIDIA GeForce RTX 3090\n\nNvidia driver version: 580.159.03\ncuDNN version: Could not collect\nIs XPU available: False\nHIP runtime version: N/A\nMIOpen runtime version: N/A\nIs XNNPACK available: True\nCaching allocator config: N/A\n\nCPU:\nArchitecture:                            x86_64\nCPU op-mode(s):                          32-bit, 64-bit\nAddress sizes:                           43 bits physical, 48 bits virtual\nByte Order:                              Little Endian\nCPU(s):                                  48\nOn-line CPU(s) list:                     0-47\nVendor ID:                               AuthenticAMD\nModel name:                              AMD EPYC 7402P 24-Core Processor\nCPU family:                              23\nModel:                                   49\nThread(s) per core:                      2\nCore(s) per socket:                      24\nSocket(s):                               1\nStepping:                                0\nFrequency boost:                         enabled\nCPU max MHz:                             3350.8049\nCPU min MHz:                             1500.0000\nBogoMIPS:                                5599.58\nFlags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sev sev_es\nVirtualization:                          AMD-V\nL1d cache:                               768 KiB (24 instances)\nL1i cache:                               768 KiB (24 instances)\nL2 cache:                                12 MiB (24 instances)\nL3 cache:                                128 MiB (8 instances)\nNUMA node(s):                            1\nNUMA node0 CPU(s):                       0-47\nVulnerability Gather data sampling:      Not affected\nVulnerability Ghostwrite:                Not affected\nVulnerability Indirect target selection: Not affected\nVulnerability Itlb multihit:             Not affected\nVulnerability L1tf:                      Not affected\nVulnerability Mds:                       Not affected\nVulnerability Meltdown:                  Not affected\nVulnerability Mmio stale data:           Not affected\nVulnerability Old microcode:             Not affected\nVulnerability Reg file data sampling:    Not affected\nVulnerability Retbleed:                  Mitigation; untrained return thunk; SMT enabled with STIBP protection\nVulnerability Spec rstack overflow:      Mitigation; Safe RET\nVulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl\nVulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization\nVulnerability Spectre v2:                Mitigation; Retpolines; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected\nVulnerability Srbds:                     Not affected\nVulnerability Tsa:                       Not affected\nVulnerability Tsx async abort:           Not affected\nVulnerability Vmscape:                   Mitigation; IBPB before exit to userspace\n\nVersions of relevant libraries:\n[pip3] Could not collect\n[conda] Could not collect",
+  "transformers_version": "4.57.6",
+  "lm_eval_version": "0.4.12",
+  "upper_git_hash": null,
+  "tokenizer_pad_token": [
+    "<pad>",
+    "50280"
+  ],
+  "tokenizer_eos_token": [
+    "|||IP_ADDRESS|||",
+    "50279"
+  ],
+  "tokenizer_bos_token": [
+    "|||IP_ADDRESS|||",
+    "50279"
+  ],
+  "eot_token_id": 50279,
+  "max_length": 4096,
+  "task_hashes": {},
+  "model_source": "hf",
+  "model_name": "allenai/OLMoE-1B-7B-0125-Instruct",
+  "model_name_sanitized": "allenai__OLMoE-1B-7B-0125-Instruct",
+  "system_instruction": null,
+  "system_instruction_sha": null,
+  "fewshot_as_multiturn": null,
+  "chat_template": null,
+  "chat_template_sha": null,
+  "total_evaluation_time_seconds": "1356.5779137950012"
+}

evals/inst_math_keep75/outputs__pruned__glean-0125inst-math-keep75/results_2026-07-10T14-48-47.027559.json

@@ -0,0 +1,205 @@
+{
+  "results": {
+    "gsm8k_cot": {
+      "name": "gsm8k_cot",
+      "alias": "gsm8k_cot",
+      "sample_len": 250,
+      "exact_match,strict-match": 0.704,
+      "exact_match_stderr,strict-match": 0.028928939388379624,
+      "exact_match,flexible-extract": 0.704,
+      "exact_match_stderr,flexible-extract": 0.028928939388379624
+    }
+  },
+  "group_subtasks": {},
+  "configs": {
+    "gsm8k_cot": {
+      "task": "gsm8k_cot",
+      "dataset_path": "openai/gsm8k",
+      "dataset_name": "main",
+      "test_split": "test",
+      "doc_to_text": "Q: {{question}}\nA:",
+      "doc_to_target": "{{answer.split('####')[-1].strip() if answer is defined else target}}",
+      "unsafe_code": false,
+      "description": "",
+      "target_delimiter": " ",
+      "fewshot_delimiter": "\n\n",
+      "fewshot_config": {
+        "sampler": "first_n",
+        "split": null,
+        "process_docs": null,
+        "fewshot_indices": null,
+        "samples": [
+          {
+            "question": "There are 15 trees in the grove. Grove workers will plant trees in the grove today. After they are done, there will be 21 trees. How many trees did the grove workers plant today?",
+            "target": "There are 15 trees originally. Then there were 21 trees after some more were planted. So there must have been 21 - 15 = 6. The answer is 6."
+          },
+          {
+            "question": "If there are 3 cars in the parking lot and 2 more cars arrive, how many cars are in the parking lot?",
+            "target": "There are originally 3 cars. 2 more cars arrive. 3 + 2 = 5. The answer is 5."
+          },
+          {
+            "question": "Leah had 32 chocolates and her sister had 42. If they ate 35, how many pieces do they have left in total?",
+            "target": "Originally, Leah had 32 chocolates. Her sister had 42. So in total they had 32 + 42 = 74. After eating 35, they had 74 - 35 = 39. The answer is 39."
+          },
+          {
+            "question": "Jason had 20 lollipops. He gave Denny some lollipops. Now Jason has 12 lollipops. How many lollipops did Jason give to Denny?",
+            "target": "Jason started with 20 lollipops. Then he had 12 after giving some to Denny. So he gave Denny 20 - 12 = 8. The answer is 8."
+          },
+          {
+            "question": "Shawn has five toys. For Christmas, he got two toys each from his mom and dad. How many toys does he have now?",
+            "target": "Shawn started with 5 toys. If he got 2 toys each from his mom and dad, then that is 4 more toys. 5 + 4 = 9. The answer is 9."
+          },
+          {
+            "question": "There were nine computers in the server room. Five more computers were installed each day, from monday to thursday. How many computers are now in the server room?",
+            "target": "There were originally 9 computers. For each of 4 days, 5 more computers were added. So 5 * 4 = 20 computers were added. 9 + 20 is 29. The answer is 29."
+          },
+          {
+            "question": "Michael had 58 golf balls. On tuesday, he lost 23 golf balls. On wednesday, he lost 2 more. How many golf balls did he have at the end of wednesday?",
+            "target": "Michael started with 58 golf balls. After losing 23 on tuesday, he had 58 - 23 = 35. After losing 2 more, he had 35 - 2 = 33 golf balls. The answer is 33."
+          },
+          {
+            "question": "Olivia has $23. She bought five bagels for $3 each. How much money does she have left?",
+            "target": "Olivia had 23 dollars. 5 bagels for 3 dollars each will be 5 x 3 = 15 dollars. So she has 23 - 15 dollars left. 23 - 15 is 8. The answer is 8."
+          }
+        ],
+        "doc_to_text": "Q: {{question}}\nA:",
+        "doc_to_choice": null,
+        "doc_to_target": "{{answer.split('####')[-1].strip() if answer is defined else target}}",
+        "gen_prefix": null,
+        "fewshot_delimiter": "\n\n",
+        "target_delimiter": " "
+      },
+      "num_fewshot": 8,
+      "metric_list": [
+        {
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "ignore_case": true,
+          "ignore_punctuation": false,
+          "metric": "exact_match",
+          "regexes_to_ignore": [
+            ",",
+            "\\$",
+            "(?s).*#### ",
+            "\\.$"
+          ]
+        }
+      ],
+      "output_type": "generate_until",
+      "generation_kwargs": {
+        "do_sample": false,
+        "until": [
+          "Q:",
+          "</s>",
+          "<|im_end|>"
+        ]
+      },
+      "repeats": 1,
+      "filter_list": [
+        {
+          "filter": [
+            {
+              "function": "regex",
+              "regex_pattern": "The answer is (\\-?[0-9\\.\\,]+)."
+            },
+            {
+              "function": "take_first"
+            }
+          ],
+          "name": "strict-match"
+        },
+        {
+          "filter": [
+            {
+              "function": "regex",
+              "group_select": -1,
+              "regex_pattern": "(-?[$0-9.,]{2,})|(-?[0-9]+)"
+            },
+            {
+              "function": "take_first"
+            }
+          ],
+          "name": "flexible-extract"
+        }
+      ],
+      "should_decontaminate": false,
+      "metadata": {
+        "version": 3.0,
+        "pretrained": "outputs/pruned/glean-0125inst-math-keep75",
+        "dtype": "bfloat16",
+        "trust_remote_code": true,
+        "config_source": "/home/henry/Documents/PythonProjects/variable-reap/.venv/lib/python3.12/site-packages/lm_eval/tasks/gsm8k/gsm8k-cot.yaml"
+      }
+    }
+  },
+  "versions": {
+    "gsm8k_cot": 3.0
+  },
+  "n-shot": {
+    "gsm8k_cot": 8
+  },
+  "higher_is_better": {
+    "gsm8k_cot": {
+      "exact_match": true
+    }
+  },
+  "n-samples": {
+    "gsm8k_cot": {
+      "original": 1319,
+      "effective": 250
+    }
+  },
+  "config": {
+    "model": "hf",
+    "model_args": {
+      "pretrained": "outputs/pruned/glean-0125inst-math-keep75",
+      "dtype": "bfloat16",
+      "trust_remote_code": true
+    },
+    "model_num_parameters": 5308368896,
+    "model_dtype": "torch.bfloat16",
+    "model_revision": "main",
+    "model_sha": "",
+    "batch_size": "auto",
+    "batch_sizes": [],
+    "device": "cuda:0",
+    "use_cache": null,
+    "limit": 250.0,
+    "bootstrap_iters": 100000,
+    "gen_kwargs": {},
+    "random_seed": 0,
+    "numpy_seed": 1234,
+    "torch_seed": 1234,
+    "fewshot_seed": 1234
+  },
+  "git_hash": "8e8c393",
+  "date": 1783718945.4797623,
+  "pretty_env_info": "PyTorch version: 2.12.1+cu130\nIs debug build: False\nCUDA used to build PyTorch: 13.0\nROCM used to build PyTorch: N/A\n\nOS: Pop!_OS 22.04 LTS (x86_64)\nGCC version: (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0\nClang version: Could not collect\nCMake version: version 3.22.1\nLibc version: glibc-2.35\n\nPython version: 3.12.12 (main, Oct 31 2025, 23:02:31) [Clang 21.1.4 ] (64-bit runtime)\nPython platform: Linux-7.0.11-76070011-generic-x86_64-with-glibc2.35\nIs CUDA available: True\nCUDA runtime version: 12.8.93\nCUDA_MODULE_LOADING set to: \nGPU models and configuration: \nGPU 0: NVIDIA GeForce RTX 3090\nGPU 1: NVIDIA GeForce RTX 3090\nGPU 2: NVIDIA GeForce RTX 3090\n\nNvidia driver version: 580.159.03\ncuDNN version: Could not collect\nIs XPU available: False\nHIP runtime version: N/A\nMIOpen runtime version: N/A\nIs XNNPACK available: True\nCaching allocator config: N/A\n\nCPU:\nArchitecture:                            x86_64\nCPU op-mode(s):                          32-bit, 64-bit\nAddress sizes:                           43 bits physical, 48 bits virtual\nByte Order:                              Little Endian\nCPU(s):                                  48\nOn-line CPU(s) list:                     0-47\nVendor ID:                               AuthenticAMD\nModel name:                              AMD EPYC 7402P 24-Core Processor\nCPU family:                              23\nModel:                                   49\nThread(s) per core:                      2\nCore(s) per socket:                      24\nSocket(s):                               1\nStepping:                                0\nFrequency boost:                         enabled\nCPU max MHz:                             3350.8049\nCPU min MHz:                             1500.0000\nBogoMIPS:                                5599.58\nFlags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sev sev_es\nVirtualization:                          AMD-V\nL1d cache:                               768 KiB (24 instances)\nL1i cache:                               768 KiB (24 instances)\nL2 cache:                                12 MiB (24 instances)\nL3 cache:                                128 MiB (8 instances)\nNUMA node(s):                            1\nNUMA node0 CPU(s):                       0-47\nVulnerability Gather data sampling:      Not affected\nVulnerability Ghostwrite:                Not affected\nVulnerability Indirect target selection: Not affected\nVulnerability Itlb multihit:             Not affected\nVulnerability L1tf:                      Not affected\nVulnerability Mds:                       Not affected\nVulnerability Meltdown:                  Not affected\nVulnerability Mmio stale data:           Not affected\nVulnerability Old microcode:             Not affected\nVulnerability Reg file data sampling:    Not affected\nVulnerability Retbleed:                  Mitigation; untrained return thunk; SMT enabled with STIBP protection\nVulnerability Spec rstack overflow:      Mitigation; Safe RET\nVulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl\nVulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization\nVulnerability Spectre v2:                Mitigation; Retpolines; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected\nVulnerability Srbds:                     Not affected\nVulnerability Tsa:                       Not affected\nVulnerability Tsx async abort:           Not affected\nVulnerability Vmscape:                   Mitigation; IBPB before exit to userspace\n\nVersions of relevant libraries:\n[pip3] Could not collect\n[conda] Could not collect",
+  "transformers_version": "4.57.6",
+  "lm_eval_version": "0.4.12",
+  "upper_git_hash": null,
+  "tokenizer_pad_token": [
+    "<pad>",
+    "50280"
+  ],
+  "tokenizer_eos_token": [
+    "|||IP_ADDRESS|||",
+    "50279"
+  ],
+  "tokenizer_bos_token": [
+    "|||IP_ADDRESS|||",
+    "50279"
+  ],
+  "eot_token_id": 50279,
+  "max_length": 4096,
+  "task_hashes": {},
+  "model_source": "hf",
+  "model_name": "outputs/pruned/glean-0125inst-math-keep75",
+  "model_name_sanitized": "outputs__pruned__glean-0125inst-math-keep75",
+  "system_instruction": null,
+  "system_instruction_sha": null,
+  "fewshot_as_multiturn": null,
+  "chat_template": null,
+  "chat_template_sha": null,
+  "total_evaluation_time_seconds": "1187.9683221830019"
+}

evals/math_unhealed/glean_winnow-olmoe-math-keep25/lm_eval.log

@@ -0,0 +1,1040 @@
+2026-08-10:14:55:50 INFO     [_cli.run:388] Selected Tasks: ['gsm8k_cot_zeroshot', 'minerva_math500', 'ifeval']
+2026-08-10:14:55:51 INFO     [evaluator:214] Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+2026-08-10:14:55:51 WARNING  [evaluator:226] generation_kwargs: {'max_gen_toks': 1280} specified through cli, these settings will update set parameters in yaml tasks. Ensure 'do_sample=True' for non-greedy decoding!
+2026-08-10:14:55:51 INFO     [evaluator:239] Initializing local-chat-completions model, with arguments: {'model': 'student', 'base_url': 'http://127.0.0.1:8522/v1/chat/completions', 'num_concurrent': 48, 'tokenized_requests': False, 'max_retries': 3}
+2026-08-10:14:55:51 INFO     [models.api_models:179] Using max length 2048 - 1
+2026-08-10:14:55:51 INFO     [models.api_models:200] Using tokenizer None
+2026-08-10:14:55:57 INFO     [evaluator_utils:446] Selected tasks:
+2026-08-10:14:55:57 INFO     [evaluator_utils:480] Task: gsm8k_cot_zeroshot (gsm8k/gsm8k-cot-zeroshot.yaml)
+2026-08-10:14:55:57 INFO     [evaluator_utils:480] Task: ifeval (ifeval/ifeval.yaml)
+2026-08-10:14:55:57 INFO     [evaluator_utils:480] Task: minerva_math500 (minerva_math/minerva_math500.yaml)
+2026-08-10:14:55:57 INFO     [evaluator:314] gsm8k_cot_zeroshot: Using gen_kwargs: {'until': ['Q:', '</s>', '<|im_end|>'], 'do_sample': False, 'max_gen_toks': 1280}
+2026-08-10:14:55:57 INFO     [evaluator:314] minerva_math500: Using gen_kwargs: {'until': ['Problem:'], 'do_sample': False, 'temperature': 0.0, 'max_gen_toks': 1280}
+2026-08-10:14:55:57 INFO     [evaluator:314] ifeval: Using gen_kwargs: {'until': [], 'do_sample': False, 'temperature': 0.0, 'max_gen_toks': 1280}
+2026-08-10:14:55:57 INFO     [api.task:312] Building contexts for gsm8k_cot_zeroshot on rank 0...
+
  0%|          | 0/1319 [00:00<?, ?it/s]
 14%|█▍        | 188/1319 [00:00<00:00, 1878.61it/s]
 29%|██▉       | 382/1319 [00:00<00:00, 1910.88it/s]
 44%|████▎     | 577/1319 [00:00<00:00, 1926.91it/s]
 58%|█████▊    | 770/1319 [00:00<00:00, 1924.97it/s]
 73%|███████▎  | 965/1319 [00:00<00:00, 1932.03it/s]
 88%|████████▊ | 1159/1319 [00:00<00:00, 1933.00it/s]
+2026-08-10:14:55:58 INFO     [api.task:312] Building contexts for minerva_math500 on rank 0...
+
  0%|          | 0/500 [00:00<?, ?it/s]
  8%|▊         | 42/500 [00:00<00:01, 410.44it/s]
 17%|█▋        | 85/500 [00:00<00:00, 416.74it/s]
 26%|██▌       | 128/500 [00:00<00:00, 419.45it/s]
 34%|███▍      | 171/500 [00:00<00:00, 420.84it/s]
 43%|████▎     | 214/500 [00:00<00:00, 421.46it/s]
 51%|█████▏    | 257/500 [00:00<00:00, 421.80it/s]
 60%|██████    | 300/500 [00:00<00:00, 422.24it/s]
 69%|██████▊   | 343/500 [00:00<00:00, 422.50it/s]
 77%|███████▋  | 386/500 [00:00<00:00, 423.41it/s]
 86%|████████▌ | 429/500 [00:01<00:00, 424.11it/s]
 94%|█████████▍| 472/500 [00:01<00:00, 424.52it/s]
+2026-08-10:14:55:59 INFO     [api.task:312] Building contexts for ifeval on rank 0...
+
  0%|          | 0/541 [00:00<?, ?it/s]
+2026-08-10:14:55:59 INFO     [evaluator:585] Running generate_until requests
+2026-08-10:14:55:59 INFO     [models.api_models:747] Tokenized requests are disabled. Context + generation length is not checked.
+
+
+
+2026-08-10:15:10:48 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text      @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ due to No features in text.
+2026-08-10:15:10:48 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text      @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ due to No features in text.
+2026-08-10:15:10:48 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text 
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+(@)
+@(@)
+@(@)
+(@)
+(=)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+(@)
+@(@)
+@(@)
+(=)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+(=)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+(=)
+@(@)
+@(@)
+(=)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+(=)
+@(@)
+@(@)
+(=)
+@(@)
+@(@)
+(=)
+@(@)
+@(@)
+(=)
+@(@)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)   (=)
+(=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)    due to No features in text.
+2026-08-10:15:10:48 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text 
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+(@)
+@(@)
+@(@)
+(@)
+(=)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+(@)
+@(@)
+@(@)
+(=)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+(=)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+(=)
+@(@)
+@(@)
+(=)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+@(@)
+(=)
+@(@)
+@(@)
+(=)
+@(@)
+@(@)
+(=)
+@(@)
+@(@)
+(=)
+@(@)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)
+(=)   (=)
+(=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)   (=)    due to No features in text.
+2026-08-10:15:10:48 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ due to No features in text.
+2026-08-10:15:10:48 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ due to No features in text.
+2026-08-10:15:10:49 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text 
+@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ due to No features in text.
+2026-08-10:15:10:49 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text 
+@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ due to No features in text.
+2026-08-10:15:10:49 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text **@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@ due to No features in text.
+2026-08-10:15:10:49 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text **@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@**@ due to No features in text.
+2026-08-10:15:10:49 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text <@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ due to No features in text.
+2026-08-10:15:10:49 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text <@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ due to No features in text.
+2026-08-10:15:10:49 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text >@
+
+>@@
+
+>@
+
+>@
+
+>@
+
+>@
+
+>@
+
+>@
+
+>@
+
+>@
+
+>@
+
+>@
+
+@
+
+@
+
+@
+
+@
+
+@
+
+@
+
+
+
+@
+
+
+
+@
+
+
+
+
+
+@
+
+
+
+@
+
+
+
+@
+
+
+
+@
+
+
+
+@
+
+@
+
+@
+
+
+
+@
+
+@
+
+@
+
+@
+
+
+
+@
+
+@
+
+
+
+@
+
+
+
+@
+
+@
+
+@
+
+
+
+@
+
+
+
+@
+
+@
+
+@
+
+@
+
+@
+
+
+
+@
+
+@
+
+@
+
+@
+
+@
+
+@
+
+@
+
+@
+
+
+
+@
+
+
+
+@
+
+@
+
+@
+
+@
+
+@
+
+@
+
+@
+
+@
+
+@
+
+
+
+@
+
+@
+
+@     @ 
+
+
+
+@     @ 
+
+
+
+@     @ 
+
+@     @     @     @     @     @     @     @     @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ 
+
+  @ @ ####   @ 
+
+  @ 
+
+  @ 
+
+  @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ due to No features in text.
+2026-08-10:15:10:49 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text @^@^2
+@^2`@'@'^'@'`'^'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`' due to No features in text.
+2026-08-10:15:10:49 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text @^@^2
+@^2`@'@'^'@'`'^'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`'`' due to No features in text.
+2026-08-10:15:10:49 ERROR    [tasks.ifeval.instructions:189] Unable to detect language for text @
+
+     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @     @ due to No features in text.
+2026-08-10:15:10:51 INFO     [loggers.evaluation_tracker:247] Saving results aggregated
+2026-08-10:15:10:51 INFO     [loggers.evaluation_tracker:119] Saving per-task samples to outputs/evals/math_unhealed/glean_winnow-olmoe-math-keep25/student/*.jsonl
+local-chat-completions ({'model': 'student', 'base_url': 'http://127.0.0.1:8522/v1/chat/completions', 'num_concurrent': 48, 'tokenized_requests': False, 'max_retries': 3}), gen_kwargs: ({'max_gen_toks': 1280}), limit: None, num_fewshot: None, batch_size: 1
+|      Tasks       |Version|     Filter     |n-shot|        Metric         |   |Value |   |Stderr|
+|------------------|------:|----------------|-----:|-----------------------|---|-----:|---|------|
+|gsm8k_cot_zeroshot|      3|flexible-extract|     0|exact_match            |↑  |0.0136|±  |0.0032|
+|                  |       |strict-match    |     0|exact_match            |↑  |0.0015|±  |0.0011|
+|ifeval            |      4|none            |     0|inst_level_loose_acc   |↑  |0.2746|±  |   N/A|
+|                  |       |none            |     0|inst_level_strict_acc  |↑  |0.2638|±  |   N/A|
+|                  |       |none            |     0|prompt_level_loose_acc |↑  |0.1627|±  |0.0159|
+|                  |       |none            |     0|prompt_level_strict_acc|↑  |0.1590|±  |0.0157|
+|minerva_math500   |      3|none            |     4|exact_match            |↑  |0.0360|±  |0.0083|
+|                  |       |none            |     4|math_verify            |↑  |0.0520|±  |0.0099|
+

evals/math_unhealed/glean_winnow-olmoe-math-keep25/lm_eval_code.log

@@ -0,0 +1,27 @@
+2026-08-10:15:11:01 INFO     [_cli.run:388] Selected Tasks: ['humaneval', 'mbpp']
+2026-08-10:15:11:02 INFO     [evaluator:214] Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+2026-08-10:15:11:02 INFO     [evaluator:239] Initializing local-completions model, with arguments: {'model': 'student', 'base_url': 'http://127.0.0.1:8522/v1/completions', 'tokenizer': 'outputs/release/unhealed/winnow-olmoe-math-keep25', 'num_concurrent': 48, 'tokenized_requests': False, 'max_retries': 3}
+2026-08-10:15:11:02 INFO     [models.openai_completions:42] Remote tokenizer not supported. Using huggingface tokenizer backend.
+2026-08-10:15:11:02 INFO     [models.api_models:179] Using max length 2048 - 1
+2026-08-10:15:11:02 INFO     [models.api_models:200] Using tokenizer huggingface
+2026-08-10:15:11:09 INFO     [evaluator_utils:446] Selected tasks:
+2026-08-10:15:11:09 INFO     [evaluator_utils:480] Task: humaneval (humaneval/humaneval.yaml)
+2026-08-10:15:11:09 INFO     [evaluator_utils:480] Task: mbpp (mbpp/mbpp.yaml)
+2026-08-10:15:11:09 INFO     [evaluator:314] humaneval: Using gen_kwargs: {'until': ['\nclass', '\ndef', '\n#', '\nif', '\nprint'], 'max_gen_toks': 1024, 'do_sample': False}
+2026-08-10:15:11:09 INFO     [evaluator:314] mbpp: Using gen_kwargs: {'until': ['[DONE]'], 'do_sample': False}
+2026-08-10:15:11:09 INFO     [api.task:312] Building contexts for humaneval on rank 0...
+
  0%|          | 0/164 [00:00<?, ?it/s]
 87%|████████▋ | 143/164 [00:00<00:00, 1424.03it/s]
+2026-08-10:15:11:09 INFO     [api.task:312] Building contexts for mbpp on rank 0...
+
  0%|          | 0/500 [00:00<?, ?it/s]
  3%|▎         | 17/500 [00:00<00:02, 161.09it/s]
  7%|▋         | 34/500 [00:00<00:02, 162.15it/s]
 10%|█         | 51/500 [00:00<00:02, 162.69it/s]
 14%|█▎        | 68/500 [00:00<00:02, 163.05it/s]
 17%|█▋        | 85/500 [00:00<00:02, 163.24it/s]
 20%|██        | 102/500 [00:00<00:02, 163.62it/s]
 24%|██▍       | 119/500 [00:00<00:02, 163.27it/s]
 27%|██▋       | 136/500 [00:00<00:02, 163.62it/s]
 31%|███       | 153/500 [00:00<00:02, 163.92it/s]
 34%|███▍      | 170/500 [00:01<00:02, 164.24it/s]
 37%|███▋      | 187/500 [00:01<00:01, 164.54it/s]
 41%|████      | 204/500 [00:01<00:01, 164.60it/s]
 44%|████▍     | 221/500 [00:01<00:01, 164.74it/s]
 48%|████▊     | 238/500 [00:01<00:01, 164.87it/s]
 51%|█████     | 255/500 [00:01<00:01, 164.98it/s]
 54%|█████▍    | 272/500 [00:01<00:01, 165.22it/s]
 58%|█████▊    | 289/500 [00:01<00:01, 165.21it/s]
 61%|██████    | 306/500 [00:01<00:01, 165.24it/s]
 65%|██████▍   | 323/500 [00:01<00:01, 165.31it/s]
 68%|██████▊   | 340/500 [00:02<00:00, 165.41it/s]
 71%|███████▏  | 357/500 [00:02<00:00, 165.45it/s]
 75%|███████▍  | 374/500 [00:02<00:00, 165.59it/s]
 78%|███████▊  | 391/500 [00:02<00:00, 165.55it/s]
 82%|████████▏ | 408/500 [00:02<00:00, 165.63it/s]
 85%|████████▌ | 425/500 [00:02<00:00, 165.67it/s]
 88%|████████▊ | 442/500 [00:02<00:00, 165.81it/s]
 92%|█████████▏| 459/500 [00:02<00:00, 165.79it/s]
 95%|█████████▌| 476/500 [00:02<00:00, 165.72it/s]
 99%|█████████▊| 493/500 [00:02<00:00, 165.94it/s]
+2026-08-10:15:11:12 INFO     [evaluator:585] Running generate_until requests
+2026-08-10:15:11:12 INFO     [models.api_models:747] Tokenized requests are disabled. Context + generation length is not checked.
+
+
+2026-08-10:15:14:49 INFO     [loggers.evaluation_tracker:247] Saving results aggregated
+2026-08-10:15:14:49 INFO     [loggers.evaluation_tracker:119] Saving per-task samples to outputs/evals/math_unhealed/glean_winnow-olmoe-math-keep25/student/*.jsonl
+local-completions ({'model': 'student', 'base_url': 'http://127.0.0.1:8522/v1/completions', 'tokenizer': 'outputs/release/unhealed/winnow-olmoe-math-keep25', 'num_concurrent': 48, 'tokenized_requests': False, 'max_retries': 3}), gen_kwargs: ({}), limit: None, num_fewshot: None, batch_size: 1
+|  Tasks  |Version|  Filter   |n-shot| Metric  |   |Value|   |Stderr|
+|---------|------:|-----------|-----:|---------|---|----:|---|-----:|
+|humaneval|      1|create_test|     0|pass@1   |↑  |    0|±  |     0|
+|mbpp     |      1|none       |     3|pass_at_1|↑  |    0|±  |     0|
+

evals/math_unhealed/glean_winnow-olmoe-math-keep50/lm_eval.log

@@ -0,0 +1,38 @@
+2026-08-10:14:55:50 INFO     [_cli.run:388] Selected Tasks: ['gsm8k_cot_zeroshot', 'minerva_math500', 'ifeval']
+2026-08-10:14:55:51 INFO     [evaluator:214] Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+2026-08-10:14:55:51 WARNING  [evaluator:226] generation_kwargs: {'max_gen_toks': 1280} specified through cli, these settings will update set parameters in yaml tasks. Ensure 'do_sample=True' for non-greedy decoding!
+2026-08-10:14:55:51 INFO     [evaluator:239] Initializing local-chat-completions model, with arguments: {'model': 'student', 'base_url': 'http://127.0.0.1:8521/v1/chat/completions', 'num_concurrent': 48, 'tokenized_requests': False, 'max_retries': 3}
+2026-08-10:14:55:51 INFO     [models.api_models:179] Using max length 2048 - 1
+2026-08-10:14:55:51 INFO     [models.api_models:200] Using tokenizer None
+2026-08-10:14:55:57 INFO     [evaluator_utils:446] Selected tasks:
+2026-08-10:14:55:57 INFO     [evaluator_utils:480] Task: gsm8k_cot_zeroshot (gsm8k/gsm8k-cot-zeroshot.yaml)
+2026-08-10:14:55:57 INFO     [evaluator_utils:480] Task: ifeval (ifeval/ifeval.yaml)
+2026-08-10:14:55:57 INFO     [evaluator_utils:480] Task: minerva_math500 (minerva_math/minerva_math500.yaml)
+2026-08-10:14:55:57 INFO     [evaluator:314] gsm8k_cot_zeroshot: Using gen_kwargs: {'until': ['Q:', '</s>', '<|im_end|>'], 'do_sample': False, 'max_gen_toks': 1280}
+2026-08-10:14:55:57 INFO     [evaluator:314] minerva_math500: Using gen_kwargs: {'until': ['Problem:'], 'do_sample': False, 'temperature': 0.0, 'max_gen_toks': 1280}
+2026-08-10:14:55:57 INFO     [evaluator:314] ifeval: Using gen_kwargs: {'until': [], 'do_sample': False, 'temperature': 0.0, 'max_gen_toks': 1280}
+2026-08-10:14:55:57 INFO     [api.task:312] Building contexts for gsm8k_cot_zeroshot on rank 0...
+
  0%|          | 0/1319 [00:00<?, ?it/s]
 14%|█▍        | 191/1319 [00:00<00:00, 1902.21it/s]
 29%|██▉       | 385/1319 [00:00<00:00, 1923.43it/s]
 44%|████▍     | 581/1319 [00:00<00:00, 1936.99it/s]
 59%|█████▉    | 777/1319 [00:00<00:00, 1942.54it/s]
 74%|███████▍  | 973/1319 [00:00<00:00, 1945.81it/s]
 89%|████████▊ | 1168/1319 [00:00<00:00, 1939.97it/s]
+2026-08-10:14:55:58 INFO     [api.task:312] Building contexts for minerva_math500 on rank 0...
+
  0%|          | 0/500 [00:00<?, ?it/s]
  8%|▊         | 42/500 [00:00<00:01, 411.34it/s]
 17%|█▋        | 84/500 [00:00<00:00, 416.06it/s]
 25%|██▌       | 127/500 [00:00<00:00, 418.71it/s]
 34%|███▍      | 170/500 [00:00<00:00, 421.00it/s]
 43%|████▎     | 213/500 [00:00<00:00, 422.35it/s]
 51%|█████     | 256/500 [00:00<00:00, 424.21it/s]
 60%|█████▉    | 299/500 [00:00<00:00, 424.80it/s]
 68%|██████▊   | 342/500 [00:00<00:00, 424.12it/s]
 77%|███████▋  | 385/500 [00:00<00:00, 424.83it/s]
 86%|████████▌ | 428/500 [00:01<00:00, 424.33it/s]
 94%|█████████▍| 471/500 [00:01<00:00, 425.60it/s]
+2026-08-10:14:55:59 INFO     [api.task:312] Building contexts for ifeval on rank 0...
+
  0%|          | 0/541 [00:00<?, ?it/s]
+2026-08-10:14:55:59 INFO     [evaluator:585] Running generate_until requests
+2026-08-10:14:55:59 INFO     [models.api_models:747] Tokenized requests are disabled. Context + generation length is not checked.
+
+
+
+2026-08-10:15:00:52 INFO     [loggers.evaluation_tracker:247] Saving results aggregated
+2026-08-10:15:00:52 INFO     [loggers.evaluation_tracker:119] Saving per-task samples to outputs/evals/math_unhealed/glean_winnow-olmoe-math-keep50/student/*.jsonl
+local-chat-completions ({'model': 'student', 'base_url': 'http://127.0.0.1:8521/v1/chat/completions', 'num_concurrent': 48, 'tokenized_requests': False, 'max_retries': 3}), gen_kwargs: ({'max_gen_toks': 1280}), limit: None, num_fewshot: None, batch_size: 1
+|      Tasks       |Version|     Filter     |n-shot|        Metric         |   |Value |   |Stderr|
+|------------------|------:|----------------|-----:|-----------------------|---|-----:|---|------|
+|gsm8k_cot_zeroshot|      3|flexible-extract|     0|exact_match            |↑  |0.4405|±  |0.0137|
+|                  |       |strict-match    |     0|exact_match            |↑  |0.0000|±  |     0|
+|ifeval            |      4|none            |     0|inst_level_loose_acc   |↑  |0.5420|±  |   N/A|
+|                  |       |none            |     0|inst_level_strict_acc  |↑  |0.5048|±  |   N/A|
+|                  |       |none            |     0|prompt_level_loose_acc |↑  |0.4177|±  |0.0212|
+|                  |       |none            |     0|prompt_level_strict_acc|↑  |0.3789|±  |0.0209|
+|minerva_math500   |      3|none            |     4|exact_match            |↑  |0.1760|±  |0.0170|
+|                  |       |none            |     4|math_verify            |↑  |0.1960|±  |0.0178|
+

evals/math_unhealed/glean_winnow-olmoe-math-keep50/lm_eval_code.log

@@ -0,0 +1,27 @@
+2026-08-10:15:01:01 INFO     [_cli.run:388] Selected Tasks: ['humaneval', 'mbpp']
+2026-08-10:15:01:02 INFO     [evaluator:214] Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+2026-08-10:15:01:02 INFO     [evaluator:239] Initializing local-completions model, with arguments: {'model': 'student', 'base_url': 'http://127.0.0.1:8521/v1/completions', 'tokenizer': 'outputs/release/unhealed/winnow-olmoe-math-keep50', 'num_concurrent': 48, 'tokenized_requests': False, 'max_retries': 3}
+2026-08-10:15:01:02 INFO     [models.openai_completions:42] Remote tokenizer not supported. Using huggingface tokenizer backend.
+2026-08-10:15:01:02 INFO     [models.api_models:179] Using max length 2048 - 1
+2026-08-10:15:01:02 INFO     [models.api_models:200] Using tokenizer huggingface
+2026-08-10:15:01:10 INFO     [evaluator_utils:446] Selected tasks:
+2026-08-10:15:01:10 INFO     [evaluator_utils:480] Task: humaneval (humaneval/humaneval.yaml)
+2026-08-10:15:01:10 INFO     [evaluator_utils:480] Task: mbpp (mbpp/mbpp.yaml)
+2026-08-10:15:01:10 INFO     [evaluator:314] humaneval: Using gen_kwargs: {'until': ['\nclass', '\ndef', '\n#', '\nif', '\nprint'], 'max_gen_toks': 1024, 'do_sample': False}
+2026-08-10:15:01:10 INFO     [evaluator:314] mbpp: Using gen_kwargs: {'until': ['[DONE]'], 'do_sample': False}
+2026-08-10:15:01:10 INFO     [api.task:312] Building contexts for humaneval on rank 0...
+
  0%|          | 0/164 [00:00<?, ?it/s]
 87%|████████▋ | 142/164 [00:00<00:00, 1419.11it/s]
+2026-08-10:15:01:10 INFO     [api.task:312] Building contexts for mbpp on rank 0...
+
  0%|          | 0/500 [00:00<?, ?it/s]
  3%|▎         | 17/500 [00:00<00:02, 161.93it/s]
  7%|▋         | 34/500 [00:00<00:02, 162.61it/s]
 10%|█         | 51/500 [00:00<00:02, 163.06it/s]
 14%|█▎        | 68/500 [00:00<00:02, 163.43it/s]
 17%|█▋        | 85/500 [00:00<00:02, 163.79it/s]
 20%|██        | 102/500 [00:00<00:02, 164.02it/s]
 24%|██▍       | 119/500 [00:00<00:02, 164.32it/s]
 27%|██▋       | 136/500 [00:00<00:02, 164.49it/s]
 31%|███       | 153/500 [00:00<00:02, 164.68it/s]
 34%|███▍      | 170/500 [00:01<00:02, 164.78it/s]
 37%|███▋      | 187/500 [00:01<00:01, 165.03it/s]
 41%|████      | 204/500 [00:01<00:01, 165.27it/s]
 44%|████▍     | 221/500 [00:01<00:01, 165.43it/s]
 48%|████▊     | 238/500 [00:01<00:01, 165.60it/s]
 51%|█████     | 255/500 [00:01<00:01, 165.58it/s]
 54%|█████▍    | 272/500 [00:01<00:01, 165.63it/s]
 58%|█████▊    | 289/500 [00:01<00:01, 165.63it/s]
 61%|██████    | 306/500 [00:01<00:01, 165.70it/s]
 65%|██████▍   | 323/500 [00:01<00:01, 165.80it/s]
 68%|██████▊   | 340/500 [00:02<00:00, 165.89it/s]
 71%|███████▏  | 357/500 [00:02<00:00, 165.80it/s]
 75%|███████▍  | 374/500 [00:02<00:00, 165.88it/s]
 78%|███████▊  | 391/500 [00:02<00:00, 165.74it/s]
 82%|████████▏ | 408/500 [00:02<00:00, 165.84it/s]
 85%|████████▌ | 425/500 [00:02<00:00, 165.82it/s]
 88%|████████▊ | 442/500 [00:02<00:00, 166.05it/s]
 92%|█████████▏| 459/500 [00:02<00:00, 165.99it/s]
 95%|█████████▌| 476/500 [00:02<00:00, 165.98it/s]
 99%|█████████▊| 493/500 [00:02<00:00, 166.07it/s]
+2026-08-10:15:01:13 INFO     [evaluator:585] Running generate_until requests
+2026-08-10:15:01:13 INFO     [models.api_models:747] Tokenized requests are disabled. Context + generation length is not checked.
+
+
+2026-08-10:15:05:32 INFO     [loggers.evaluation_tracker:247] Saving results aggregated
+2026-08-10:15:05:32 INFO     [loggers.evaluation_tracker:119] Saving per-task samples to outputs/evals/math_unhealed/glean_winnow-olmoe-math-keep50/student/*.jsonl
+local-completions ({'model': 'student', 'base_url': 'http://127.0.0.1:8521/v1/completions', 'tokenizer': 'outputs/release/unhealed/winnow-olmoe-math-keep50', 'num_concurrent': 48, 'tokenized_requests': False, 'max_retries': 3}), gen_kwargs: ({}), limit: None, num_fewshot: None, batch_size: 1
+|  Tasks  |Version|  Filter   |n-shot| Metric  |   |Value |   |Stderr|
+|---------|------:|-----------|-----:|---------|---|-----:|---|-----:|
+|humaneval|      1|create_test|     0|pass@1   |↑  |0.0732|±  |0.0204|
+|mbpp     |      1|none       |     3|pass_at_1|↑  |0.0900|±  |0.0128|
+