ruihan/qwen38-flash-next-dgx-spark
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Document long-context limits and host OOM at near-262K input

Browse Source
This commit is contained in:
ruihan
2026-09-18 00:03:19 +08:00
parent 48a1c9d7c4
commit f1a8964072
16 changed files with 1054 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/results/context262k/graph.jsonl
+52
View File
@@ -0,0 +1,52 @@
{"event": "STAGE_START", "label": "graph", "target_input_tokens": 65536}
{"event": "DOCUMENT_START", "label": "graph", "document": "ctx-65536-1", "document_sha256": "62dc71efdb9de440f4cf319472ece04e2fe59e3c31653d7a95ecf050549c1697", "needle_token_fractions": [0.0502, 0.4994, 0.9486]}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-65536-1", "case": "front"}
{"event": "RESULT", "label": "graph", "document": "ctx-65536-1", "target_input_tokens": 65536, "sample": 1, "case": "front", "attempt": 1, "expected": "90444", "content": "\n\n90444", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 65510, "total_tokens": 65575, "completion_tokens": 65, "completion_tokens_details": {"reasoning_tokens": 57}}, "max_tokens": 2048, "prompt_tokens_preflight": 65510, "prompt_sha256": "cae6de2523bb6e51419630c9a0def8f632e526a3dd32af08f2fd72eff9f3f2d4", "output_sha256": "302f94d12b4e252baa7baf220d83b1bcf00ab99aa2cbe0caf0f9a2f11c136667", "ttft_s": 31.3812, "first_content_s": 33.1416, "elapsed_s": 33.3135, "decode_tps_approx": 34.82, "metrics_delta": {"vllm:prefix_cache_hits_total": 0.0, "vllm:prefix_cache_queries_total": 65510.0, "vllm:num_preemptions_total": 0.0}}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-65536-1", "case": "middle"}
{"event": "RESULT", "label": "graph", "document": "ctx-65536-1", "target_input_tokens": 65536, "sample": 1, "case": "middle", "attempt": 1, "expected": "34096", "content": "\n\n34096", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 65509, "total_tokens": 65574, "completion_tokens": 65, "completion_tokens_details": {"reasoning_tokens": 57}}, "max_tokens": 2048, "prompt_tokens_preflight": 65509, "prompt_sha256": "4aed1cfbe19cb56496e9496753062083ebb9382611731804e6c867ae1025795c", "output_sha256": "9c9329a6672fc532dc4ea0795ac06c2c1719af59592c082b2b3bcad75f10cc9a", "ttft_s": 1.6313, "first_content_s": 3.4207, "elapsed_s": 3.593, "decode_tps_approx": 32.744, "metrics_delta": {"vllm:prefix_cache_hits_total": 62400.0, "vllm:prefix_cache_queries_total": 65509.0, "vllm:num_preemptions_total": 0.0}}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-65536-1", "case": "end"}
{"event": "RESULT", "label": "graph", "document": "ctx-65536-1", "target_input_tokens": 65536, "sample": 1, "case": "end", "attempt": 1, "expected": "90059", "content": "\n\n90059", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 65508, "total_tokens": 65571, "completion_tokens": 63, "completion_tokens_details": {"reasoning_tokens": 55}}, "max_tokens": 2048, "prompt_tokens_preflight": 65508, "prompt_sha256": "9415a83f1f499579d7d90bfd4adf63184b021672491f843e6fdea865fdefe9e3", "output_sha256": "f1732242e69c1661c79495c8e0f95a68dee3b8e688ec39fb928caf4fdf58c9f1", "ttft_s": 1.6231, "first_content_s": 3.2847, "elapsed_s": 3.4708, "decode_tps_approx": 33.694, "metrics_delta": {"vllm:prefix_cache_hits_total": 62400.0, "vllm:prefix_cache_queries_total": 65508.0, "vllm:num_preemptions_total": 0.0}}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-65536-1", "case": "front_repeat"}
{"event": "RESULT", "label": "graph", "document": "ctx-65536-1", "target_input_tokens": 65536, "sample": 1, "case": "front_repeat", "attempt": 1, "expected": "90444", "content": "\n\n90444", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 65510, "total_tokens": 65575, "completion_tokens": 65, "completion_tokens_details": {"reasoning_tokens": 57}}, "max_tokens": 2048, "prompt_tokens_preflight": 65510, "prompt_sha256": "cae6de2523bb6e51419630c9a0def8f632e526a3dd32af08f2fd72eff9f3f2d4", "output_sha256": "302f94d12b4e252baa7baf220d83b1bcf00ab99aa2cbe0caf0f9a2f11c136667", "ttft_s": 1.631, "first_content_s": 3.3, "elapsed_s": 3.4747, "decode_tps_approx": 36.506, "metrics_delta": {"vllm:prefix_cache_hits_total": 62400.0, "vllm:prefix_cache_queries_total": 65510.0, "vllm:num_preemptions_total": 0.0}}
{"event": "DOCUMENT_START", "label": "graph", "document": "ctx-65536-2", "document_sha256": "075f6a50b9fbbc275aaea62f2650a46838f4d19e93a543c4572c4e7acaf90988", "needle_token_fractions": [0.0502, 0.4994, 0.9486]}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-65536-2", "case": "front"}
{"event": "RESULT", "label": "graph", "document": "ctx-65536-2", "target_input_tokens": 65536, "sample": 2, "case": "front", "attempt": 1, "expected": "22735", "content": "\n\n22735", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 65510, "total_tokens": 65577, "completion_tokens": 67, "completion_tokens_details": {"reasoning_tokens": 59}}, "max_tokens": 2048, "prompt_tokens_preflight": 65510, "prompt_sha256": "a42dc20d962b640ab70d2b748464a7907983a252746f927b96edd3c822069830", "output_sha256": "800ef0d2008bd683590ca51be8004d2b067f700b5b4678dbc6838671ae14004f", "ttft_s": 31.2016, "first_content_s": 32.9539, "elapsed_s": 33.1287, "decode_tps_approx": 34.386, "metrics_delta": {"vllm:prefix_cache_hits_total": 0.0, "vllm:prefix_cache_queries_total": 65510.0, "vllm:num_preemptions_total": 0.0}}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-65536-2", "case": "middle"}
{"event": "RESULT", "label": "graph", "document": "ctx-65536-2", "target_input_tokens": 65536, "sample": 2, "case": "middle", "attempt": 1, "expected": "98413", "content": "\n\n98413", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 65509, "total_tokens": 65576, "completion_tokens": 67, "completion_tokens_details": {"reasoning_tokens": 59}}, "max_tokens": 2048, "prompt_tokens_preflight": 65509, "prompt_sha256": "3dc3b7fe628cc09e0ca8d064cbb216150736b71e2913e60610b0baf76dcc01d9", "output_sha256": "0776f36146e6ef2c6ce63b0167fbd8477a8029af20d4d422d30dad32e15cf392", "ttft_s": 1.634, "first_content_s": 3.4029, "elapsed_s": 3.5735, "decode_tps_approx": 35.588, "metrics_delta": {"vllm:prefix_cache_hits_total": 62400.0, "vllm:prefix_cache_queries_total": 65509.0, "vllm:num_preemptions_total": 0.0}}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-65536-2", "case": "end"}
{"event": "RESULT", "label": "graph", "document": "ctx-65536-2", "target_input_tokens": 65536, "sample": 2, "case": "end", "attempt": 1, "expected": "61613", "content": "\n\n61613", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 65508, "total_tokens": 65574, "completion_tokens": 66, "completion_tokens_details": {"reasoning_tokens": 58}}, "max_tokens": 2048, "prompt_tokens_preflight": 65508, "prompt_sha256": "96f91ecd7e78d9c40fa68491b95c9a7679ecd8ac7cdf8cf48bea95317e467d9a", "output_sha256": "f0d613427a49bcc6875f076a4bacab0e199cb28c02361640f1d0dd76b12cd42a", "ttft_s": 1.6339, "first_content_s": 3.3157, "elapsed_s": 3.4025, "decode_tps_approx": 36.91, "metrics_delta": {"vllm:prefix_cache_hits_total": 62400.0, "vllm:prefix_cache_queries_total": 65508.0, "vllm:num_preemptions_total": 0.0}}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-65536-2", "case": "front_repeat"}
{"event": "RESULT", "label": "graph", "document": "ctx-65536-2", "target_input_tokens": 65536, "sample": 2, "case": "front_repeat", "attempt": 1, "expected": "22735", "content": "\n\n22735", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 65510, "total_tokens": 65581, "completion_tokens": 71, "completion_tokens_details": {"reasoning_tokens": 63}}, "max_tokens": 2048, "prompt_tokens_preflight": 65510, "prompt_sha256": "a42dc20d962b640ab70d2b748464a7907983a252746f927b96edd3c822069830", "output_sha256": "f8f3af6416a04184ba903b37a60fae8d67d230abfba70767e09d494429cedd33", "ttft_s": 1.6437, "first_content_s": 3.5088, "elapsed_s": 3.6774, "decode_tps_approx": 34.565, "metrics_delta": {"vllm:prefix_cache_hits_total": 62400.0, "vllm:prefix_cache_queries_total": 65510.0, "vllm:num_preemptions_total": 0.0}}
{"event": "STAGE_COMPLETED", "label": "graph", "target_input_tokens": 65536}
{"event": "STAGE_START", "label": "graph", "target_input_tokens": 131072}
{"event": "DOCUMENT_START", "label": "graph", "document": "ctx-131072-1", "document_sha256": "81014bff67378118f3b3ceba03c745724783df5600e0327fc34d8b57616b48c1", "needle_token_fractions": [0.05, 0.4998, 0.9494]}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-131072-1", "case": "front"}
{"event": "RESULT", "label": "graph", "document": "ctx-131072-1", "target_input_tokens": 131072, "sample": 1, "case": "front", "attempt": 1, "expected": "22798", "content": "\n\n22798", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 131040, "total_tokens": 131107, "completion_tokens": 67, "completion_tokens_details": {"reasoning_tokens": 59}}, "max_tokens": 2048, "prompt_tokens_preflight": 131040, "prompt_sha256": "941a60b4a55cfbf2250885fd7af910fcee51df9a63a7bb0e5fd23eb0b2e981c9", "output_sha256": "f2e65532c3fedd9134daa4660cdd96280c3283b7ef9f4514d1b54f865d5c0c65", "ttft_s": 64.6686, "first_content_s": 66.4772, "elapsed_s": 66.5778, "decode_tps_approx": 34.72, "metrics_delta": {"vllm:prefix_cache_hits_total": 0.0, "vllm:prefix_cache_queries_total": 131040.0, "vllm:num_preemptions_total": 0.0}}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-131072-1", "case": "middle"}
{"event": "RESULT", "label": "graph", "document": "ctx-131072-1", "target_input_tokens": 131072, "sample": 1, "case": "middle", "attempt": 1, "expected": "87671", "content": "\n\n87671", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 131039, "total_tokens": 131104, "completion_tokens": 65, "completion_tokens_details": {"reasoning_tokens": 57}}, "max_tokens": 2048, "prompt_tokens_preflight": 131039, "prompt_sha256": "7289422da2bb1b74f436e7589057e2ee045b3fb9f6a9d24af6c560f344768a30", "output_sha256": "a332a7f3eb061442f33f29bb29a7727c2d71ddd9afa5bea2584bb532b5e27920", "ttft_s": 1.7696, "first_content_s": 3.5958, "elapsed_s": 3.7695, "decode_tps_approx": 32.147, "metrics_delta": {"vllm:prefix_cache_hits_total": 128000.0, "vllm:prefix_cache_queries_total": 131039.0, "vllm:num_preemptions_total": 0.0}}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-131072-1", "case": "end"}
{"event": "RESULT", "label": "graph", "document": "ctx-131072-1", "target_input_tokens": 131072, "sample": 1, "case": "end", "attempt": 1, "expected": "66667", "content": "\n\n66667", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 131038, "total_tokens": 131106, "completion_tokens": 68, "completion_tokens_details": {"reasoning_tokens": 60}}, "max_tokens": 2048, "prompt_tokens_preflight": 131038, "prompt_sha256": "5231b28263eefbc827fce10ebfe9af3f43baf90154bfd9ddac916ca086ad1eda", "output_sha256": "b2e1d3b077c4c2780190d7705a5569da88f63ef165a81b6ad5b6f0f81ec80286", "ttft_s": 1.7634, "first_content_s": 3.669, "elapsed_s": 3.7757, "decode_tps_approx": 33.435, "metrics_delta": {"vllm:prefix_cache_hits_total": 128000.0, "vllm:prefix_cache_queries_total": 131038.0, "vllm:num_preemptions_total": 0.0}}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-131072-1", "case": "front_repeat"}
{"event": "RESULT", "label": "graph", "document": "ctx-131072-1", "target_input_tokens": 131072, "sample": 1, "case": "front_repeat", "attempt": 1, "expected": "22798", "content": "\n\n22798", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 131040, "total_tokens": 131107, "completion_tokens": 67, "completion_tokens_details": {"reasoning_tokens": 59}}, "max_tokens": 2048, "prompt_tokens_preflight": 131040, "prompt_sha256": "941a60b4a55cfbf2250885fd7af910fcee51df9a63a7bb0e5fd23eb0b2e981c9", "output_sha256": "8e8ae60c86d7472427ad9fa0f03351e31fc544c601d3a20bcf41e6e7902cc5df", "ttft_s": 1.7666, "first_content_s": 3.4995, "elapsed_s": 3.6726, "decode_tps_approx": 34.776, "metrics_delta": {"vllm:prefix_cache_hits_total": 128000.0, "vllm:prefix_cache_queries_total": 131040.0, "vllm:num_preemptions_total": 0.0}}
{"event": "DOCUMENT_START", "label": "graph", "document": "ctx-131072-2", "document_sha256": "040a0cd85bb25a86dd2bb771a070e8bdc8a8c64b7de671f63e35a5a30cf2ee59", "needle_token_fractions": [0.05, 0.4998, 0.9494]}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-131072-2", "case": "front"}
{"event": "RESULT", "label": "graph", "document": "ctx-131072-2", "target_input_tokens": 131072, "sample": 2, "case": "front", "attempt": 1, "expected": "21330", "content": "\n\n21330", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 131040, "total_tokens": 131106, "completion_tokens": 66, "completion_tokens_details": {"reasoning_tokens": 58}}, "max_tokens": 2048, "prompt_tokens_preflight": 131040, "prompt_sha256": "cbbe2526ffeabd635cb332d2884e9b38cfd848fc01920eff66bb2ba53a16aa09", "output_sha256": "d0ed7218b33c233b8f19cfe2d065b167568dcb52d2c41b9006d672e9478aee7d", "ttft_s": 63.2781, "first_content_s": 65.0856, "elapsed_s": 65.2604, "decode_tps_approx": 32.929, "metrics_delta": {"vllm:prefix_cache_hits_total": 0.0, "vllm:prefix_cache_queries_total": 131040.0, "vllm:num_preemptions_total": 0.0}}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-131072-2", "case": "middle"}
{"event": "RESULT", "label": "graph", "document": "ctx-131072-2", "target_input_tokens": 131072, "sample": 2, "case": "middle", "attempt": 1, "expected": "22858", "content": "\n\n22858", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 131039, "total_tokens": 131107, "completion_tokens": 68, "completion_tokens_details": {"reasoning_tokens": 60}}, "max_tokens": 2048, "prompt_tokens_preflight": 131039, "prompt_sha256": "68107d0867abd73f9abe591916a726eaaef94e3fc5d0277566a76ea49803a4cb", "output_sha256": "f07af8e77a0089c8356262cc1a11056bae37f3d44f3fded5c0bbdb5454e74eb8", "ttft_s": 1.7672, "first_content_s": 3.5913, "elapsed_s": 3.699, "decode_tps_approx": 34.823, "metrics_delta": {"vllm:prefix_cache_hits_total": 128000.0, "vllm:prefix_cache_queries_total": 131039.0, "vllm:num_preemptions_total": 0.0}}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-131072-2", "case": "end"}
{"event": "RESULT", "label": "graph", "document": "ctx-131072-2", "target_input_tokens": 131072, "sample": 2, "case": "end", "attempt": 1, "expected": "26681", "content": "\n\n26681", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 131038, "total_tokens": 131101, "completion_tokens": 63, "completion_tokens_details": {"reasoning_tokens": 55}}, "max_tokens": 2048, "prompt_tokens_preflight": 131038, "prompt_sha256": "344c248b70b3519e8446e645d9afdfafe315ebfb7a619e70ff458d6b0cc628e1", "output_sha256": "8c75d9d6e54e6f3a5f5deb1a51d2c06f7deb0697c53e7a8a4718163490d5292e", "ttft_s": 1.7714, "first_content_s": 3.4271, "elapsed_s": 3.6039, "decode_tps_approx": 33.993, "metrics_delta": {"vllm:prefix_cache_hits_total": 128000.0, "vllm:prefix_cache_queries_total": 131038.0, "vllm:num_preemptions_total": 0.0}}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-131072-2", "case": "front_repeat"}
{"event": "RESULT", "label": "graph", "document": "ctx-131072-2", "target_input_tokens": 131072, "sample": 2, "case": "front_repeat", "attempt": 1, "expected": "21330", "content": "\n\n21330", "correct": true, "finish_reason": "stop", "usage": {"prompt_tokens": 131040, "total_tokens": 131107, "completion_tokens": 67, "completion_tokens_details": {"reasoning_tokens": 59}}, "max_tokens": 2048, "prompt_tokens_preflight": 131040, "prompt_sha256": "cbbe2526ffeabd635cb332d2884e9b38cfd848fc01920eff66bb2ba53a16aa09", "output_sha256": "e10c3e14917479637c124594ac830790a96d90ed04093d982f4729fe678123ad", "ttft_s": 1.7713, "first_content_s": 3.5814, "elapsed_s": 3.7475, "decode_tps_approx": 33.552, "metrics_delta": {"vllm:prefix_cache_hits_total": 128000.0, "vllm:prefix_cache_queries_total": 131040.0, "vllm:num_preemptions_total": 0.0}}
{"event": "STAGE_COMPLETED", "label": "graph", "target_input_tokens": 131072}
{"event": "STAGE_START", "label": "graph", "target_input_tokens": 260000}
{"event": "DOCUMENT_START", "label": "graph", "document": "ctx-260000-1", "document_sha256": "35126a00209a0d6d178dbe3aaa810912e87134bd2cc842fb9a038f890e8a1602", "needle_token_fractions": [0.05, 0.4998, 0.9497]}
{"event": "REQUEST_START", "label": "graph", "document": "ctx-260000-1", "case": "front"}
Traceback (most recent call last):
File "/tmp/context-benchmark.py", line 198, in <module>
run(args.data, args.label)
File "/tmp/context-benchmark.py", line 180, in run
result = run_case(label, fixture, case, 1)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/tmp/context-benchmark.py", line 148, in run_case
raise RuntimeError('Incomplete response or missing usage')
RuntimeError: Incomplete response or missing usage