{ "status": "untrained_runtime_probe_not_quality_or_calibration_evidence", "model": "models/gemma-4-e2b-it-4bit", "state_layers": 4, "compute_dtype": "float16", "head_config": { "input_dims": 1536, "dims": 256, "heads": 4, "query_layers": 1, "decision_layers": 2 }, "timing_scope": "warm_model_fresh_state_including_media_decode_preprocess_encoding_and_all_fields", "field_scaling": "four representative question templates repeated to the requested count", "results": [ { "state": "text", "fields": 1, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.0003517500008456409, "state_encode_seconds": 0.008788832987193018, "query_embed_seconds": 0.0004464590165298432, "trained": false, "head_seconds": 0.0009058340219780803, "total_seconds": 0.010501707991352305, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.00021275001927278936, "state_encode_seconds": 0.008485166996251792, "query_embed_seconds": 0.0002860829990822822, "trained": false, "head_seconds": 0.004243875009706244, "total_seconds": 0.013233250007033348, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.00025104201631620526, "state_encode_seconds": 0.008505375008098781, "query_embed_seconds": 0.0003567919775377959, "trained": false, "head_seconds": 0.0008325410017278045, "total_seconds": 0.009951624990208074, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.000277625018497929, "state_encode_seconds": 0.008535333996405825, "query_embed_seconds": 0.00035691598895937204, "trained": false, "head_seconds": 0.0008846669807098806, "total_seconds": 0.010061999986646697, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.0002744169905781746, "state_encode_seconds": 0.012299750000238419, "query_embed_seconds": 0.0003435419930610806, "trained": false, "head_seconds": 0.0008590829966124147, "total_seconds": 0.013783542002784088, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.0002744169905781746, "state_encode_seconds": 0.008535333996405825, "query_embed_seconds": 0.0003567919775377959, "head_seconds": 0.0008846669807098806, "total_seconds": 0.010501707991352305 }, "max_seconds": 0.013783542002784088, "under_300ms": 5 }, { "state": "text", "fields": 4, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.00029674998950213194, "state_encode_seconds": 0.008566292002797127, "query_embed_seconds": 0.0004583330010063946, "trained": false, "head_seconds": 0.0008601660083513707, "total_seconds": 0.010187459003645927, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.0002726250095292926, "state_encode_seconds": 0.008499374991515651, "query_embed_seconds": 0.00041625002631917596, "trained": false, "head_seconds": 0.000871583994012326, "total_seconds": 0.010064999980386347, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.00024358299560844898, "state_encode_seconds": 0.008250791986938566, "query_embed_seconds": 0.003285542014054954, "trained": false, "head_seconds": 0.0013534580066334456, "total_seconds": 0.013143958000000566, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.00026287499349564314, "state_encode_seconds": 0.00842654099687934, "query_embed_seconds": 0.00048424999113194644, "trained": false, "head_seconds": 0.0009105830104090273, "total_seconds": 0.010094333003507927, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.00026183301815763116, "state_encode_seconds": 0.008452916983515024, "query_embed_seconds": 0.00046304200077429414, "trained": false, "head_seconds": 0.0008629579970147461, "total_seconds": 0.010045790986623615, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.00026287499349564314, "state_encode_seconds": 0.008452916983515024, "query_embed_seconds": 0.00046304200077429414, "head_seconds": 0.000871583994012326, "total_seconds": 0.010094333003507927 }, "max_seconds": 0.013143958000000566, "under_300ms": 5 }, { "state": "text", "fields": 16, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.0002402920217718929, "state_encode_seconds": 0.0086124999797903, "query_embed_seconds": 0.0007817080186214298, "trained": false, "head_seconds": 0.0049997080059256405, "total_seconds": 0.014639084023656324, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.00028987499536015093, "state_encode_seconds": 0.008837500005029142, "query_embed_seconds": 0.0009521669999230653, "trained": false, "head_seconds": 0.0011727090168278664, "total_seconds": 0.011260208993917331, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.0002667499938979745, "state_encode_seconds": 0.008477499999571592, "query_embed_seconds": 0.0008369580027647316, "trained": false, "head_seconds": 0.0011303330247756094, "total_seconds": 0.010717583994846791, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.00026129200705327094, "state_encode_seconds": 0.008356375008588657, "query_embed_seconds": 0.004173458000877872, "trained": false, "head_seconds": 0.0011804999958258122, "total_seconds": 0.01397970801917836, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.000310958013869822, "state_encode_seconds": 0.00843591699958779, "query_embed_seconds": 0.0009015419927891344, "trained": false, "head_seconds": 0.001880750001873821, "total_seconds": 0.011536791018443182, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.0002667499938979745, "state_encode_seconds": 0.008477499999571592, "query_embed_seconds": 0.0009015419927891344, "head_seconds": 0.0011804999958258122, "total_seconds": 0.011536791018443182 }, "max_seconds": 0.014639084023656324, "under_300ms": 5 }, { "state": "text", "fields": 28, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.00023591599892824888, "state_encode_seconds": 0.008571125013986602, "query_embed_seconds": 0.005022041994379833, "trained": false, "head_seconds": 0.0015913749812170863, "total_seconds": 0.01543862500693649, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.0003165829984936863, "state_encode_seconds": 0.008780083007877693, "query_embed_seconds": 0.0012893749808426946, "trained": false, "head_seconds": 0.00558554200688377, "total_seconds": 0.015978415991412476, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.0002532500075176358, "state_encode_seconds": 0.008575792016927153, "query_embed_seconds": 0.001213207986438647, "trained": false, "head_seconds": 0.0027006669843103737, "total_seconds": 0.012749499990604818, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.000317083002300933, "state_encode_seconds": 0.010130917013157159, "query_embed_seconds": 0.0012695000041276217, "trained": false, "head_seconds": 0.0014616670086979866, "total_seconds": 0.013186708994908258, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 50, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.0003190409915987402, "state_encode_seconds": 0.010874875006265938, "query_embed_seconds": 0.0012795840157195926, "trained": false, "head_seconds": 0.0015025830070953816, "total_seconds": 0.013984874996822327, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.0003165829984936863, "state_encode_seconds": 0.008780083007877693, "query_embed_seconds": 0.0012795840157195926, "head_seconds": 0.0015913749812170863, "total_seconds": 0.013984874996822327 }, "max_seconds": 0.015978415991412476, "under_300ms": 5 }, { "state": "image", "fields": 1, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.004423208010848612, "state_encode_seconds": 0.33653075000620447, "query_embed_seconds": 0.0004351249954197556, "trained": false, "head_seconds": 0.004715082992333919, "total_seconds": 0.346112458006246, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.004197458998532966, "state_encode_seconds": 0.3318729579914361, "query_embed_seconds": 0.00045125000178813934, "trained": false, "head_seconds": 0.004745457990793511, "total_seconds": 0.3412757499900181, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.004103915998712182, "state_encode_seconds": 0.3372717919992283, "query_embed_seconds": 0.00042316701728850603, "trained": false, "head_seconds": 0.0009011670190375298, "total_seconds": 0.3427087079908233, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.004504375014221296, "state_encode_seconds": 0.48244283298845403, "query_embed_seconds": 0.0017175830143969506, "trained": false, "head_seconds": 0.0029679159924853593, "total_seconds": 0.4916505420114845, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.008557916997233406, "state_encode_seconds": 0.5831141250091605, "query_embed_seconds": 0.0005946249875705689, "trained": false, "head_seconds": 0.0011208750074729323, "total_seconds": 0.5934008340118453, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.004423208010848612, "state_encode_seconds": 0.3372717919992283, "query_embed_seconds": 0.00045125000178813934, "head_seconds": 0.0029679159924853593, "total_seconds": 0.346112458006246 }, "max_seconds": 0.5934008340118453, "under_300ms": 0 }, { "state": "image", "fields": 4, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.004069291986525059, "state_encode_seconds": 0.3392385420156643, "query_embed_seconds": 0.000559957989025861, "trained": false, "head_seconds": 0.00104674999602139, "total_seconds": 0.3449216250155587, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.004686500004027039, "state_encode_seconds": 0.33156995801255107, "query_embed_seconds": 0.0032142499985639006, "trained": false, "head_seconds": 0.0013713750231545419, "total_seconds": 0.340856334019918, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.004511208011535928, "state_encode_seconds": 0.33989699999801815, "query_embed_seconds": 0.0034782500006258488, "trained": false, "head_seconds": 0.0014392909943126142, "total_seconds": 0.34934345801593736, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.004411707981489599, "state_encode_seconds": 0.4570657080039382, "query_embed_seconds": 0.0036302920198068023, "trained": false, "head_seconds": 0.001170917006675154, "total_seconds": 0.46629141698940657, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.005538041994441301, "state_encode_seconds": 0.5729153749998659, "query_embed_seconds": 0.0005725410010199994, "trained": false, "head_seconds": 0.0010882919887080789, "total_seconds": 0.5801238749991171, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.004511208011535928, "state_encode_seconds": 0.33989699999801815, "query_embed_seconds": 0.0032142499985639006, "head_seconds": 0.001170917006675154, "total_seconds": 0.34934345801593736 }, "max_seconds": 0.5801238749991171, "under_300ms": 0 }, { "state": "image", "fields": 16, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.004117625008802861, "state_encode_seconds": 0.33643283299170434, "query_embed_seconds": 0.000992834015050903, "trained": false, "head_seconds": 0.005133708997163922, "total_seconds": 0.34668712501297705, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.00413337501231581, "state_encode_seconds": 0.33063116698758677, "query_embed_seconds": 0.004334041994297877, "trained": false, "head_seconds": 0.0012911660014651716, "total_seconds": 0.3404005410266109, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.00415841699577868, "state_encode_seconds": 0.33665308298077434, "query_embed_seconds": 0.0009698750218376517, "trained": false, "head_seconds": 0.0014100410044193268, "total_seconds": 0.34319983300520107, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.004215082997689024, "state_encode_seconds": 0.33648595900740474, "query_embed_seconds": 0.0010194159985985607, "trained": false, "head_seconds": 0.0013218340172898024, "total_seconds": 0.3430620830040425, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.004183042008662596, "state_encode_seconds": 0.34372070801327936, "query_embed_seconds": 0.0010020829795394093, "trained": false, "head_seconds": 0.0014439169899560511, "total_seconds": 0.35035925000556745, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.00415841699577868, "state_encode_seconds": 0.33648595900740474, "query_embed_seconds": 0.0010020829795394093, "head_seconds": 0.0014100410044193268, "total_seconds": 0.34319983300520107 }, "max_seconds": 0.35035925000556745, "under_300ms": 0 }, { "state": "image", "fields": 28, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.00408141702064313, "state_encode_seconds": 0.3329160419816617, "query_embed_seconds": 0.004642540996428579, "trained": false, "head_seconds": 0.0019210829923395067, "total_seconds": 0.34358479201910086, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.0043738750100601465, "state_encode_seconds": 0.3350752500118688, "query_embed_seconds": 0.0013027919922024012, "trained": false, "head_seconds": 0.0015492920065298676, "total_seconds": 0.34230887502781115, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.004092832998139784, "state_encode_seconds": 0.33910854198620655, "query_embed_seconds": 0.004882750014076009, "trained": false, "head_seconds": 0.001895624998724088, "total_seconds": 0.3499921250040643, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.004197333997581154, "state_encode_seconds": 0.3435345409961883, "query_embed_seconds": 0.0013600419915746897, "trained": false, "head_seconds": 0.0015774999919813126, "total_seconds": 0.3506769160158001, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 295, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.004053625016240403, "state_encode_seconds": 0.35448620800161734, "query_embed_seconds": 0.005586209008470178, "trained": false, "head_seconds": 0.0019304160086903721, "total_seconds": 0.3660713329736609, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.004092832998139784, "state_encode_seconds": 0.33910854198620655, "query_embed_seconds": 0.004642540996428579, "head_seconds": 0.001895624998724088, "total_seconds": 0.3499921250040643 }, "max_seconds": 0.3660713329736609, "under_300ms": 0 }, { "state": "speech", "fields": 1, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.0330117919947952, "state_encode_seconds": 0.06897991700679995, "query_embed_seconds": 0.00044470798457041383, "trained": false, "head_seconds": 0.0009053749963641167, "total_seconds": 0.10336504198494367, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.03472291701473296, "state_encode_seconds": 0.06393716597813182, "query_embed_seconds": 0.0004162499972153455, "trained": false, "head_seconds": 0.0009512079996056855, "total_seconds": 0.10006216602050699, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.0328192499873694, "state_encode_seconds": 0.0657884580141399, "query_embed_seconds": 0.000464374985313043, "trained": false, "head_seconds": 0.0013164169795345515, "total_seconds": 0.10041033299057744, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.041251457994803786, "state_encode_seconds": 0.07002408299013041, "query_embed_seconds": 0.0005837500211782753, "trained": false, "head_seconds": 0.007458249980118126, "total_seconds": 0.11934283300070092, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.10829504099092446, "state_encode_seconds": 0.08610745900659822, "query_embed_seconds": 0.0004944580141454935, "trained": false, "head_seconds": 0.001001208001980558, "total_seconds": 0.1959179590048734, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.03472291701473296, "state_encode_seconds": 0.06897991700679995, "query_embed_seconds": 0.000464374985313043, "head_seconds": 0.001001208001980558, "total_seconds": 0.10336504198494367 }, "max_seconds": 0.1959179590048734, "under_300ms": 5 }, { "state": "speech", "fields": 4, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.033792292000725865, "state_encode_seconds": 0.07191387499915436, "query_embed_seconds": 0.0005895830108784139, "trained": false, "head_seconds": 0.0009822919964790344, "total_seconds": 0.10732433298835531, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.03342108300421387, "state_encode_seconds": 0.0664535420073662, "query_embed_seconds": 0.0006965829816181213, "trained": false, "head_seconds": 0.0009517500002402812, "total_seconds": 0.1015518750064075, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.0348786250106059, "state_encode_seconds": 0.0738653339794837, "query_embed_seconds": 0.0007122080132830888, "trained": false, "head_seconds": 0.005216541001573205, "total_seconds": 0.11468895801226608, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.038284875015961006, "state_encode_seconds": 0.07466908299829811, "query_embed_seconds": 0.0036092500085942447, "trained": false, "head_seconds": 0.0016579999937675893, "total_seconds": 0.11832049998338334, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.035641707974718884, "state_encode_seconds": 0.06496925000101328, "query_embed_seconds": 0.0028525000088848174, "trained": false, "head_seconds": 0.0018704170070122927, "total_seconds": 0.1053697080060374, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.0348786250106059, "state_encode_seconds": 0.07191387499915436, "query_embed_seconds": 0.0007122080132830888, "head_seconds": 0.0016579999937675893, "total_seconds": 0.10732433298835531 }, "max_seconds": 0.11832049998338334, "under_300ms": 5 }, { "state": "speech", "fields": 16, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.03924745798576623, "state_encode_seconds": 0.08621008400223218, "query_embed_seconds": 0.001717915991321206, "trained": false, "head_seconds": 0.004675874981330708, "total_seconds": 0.13189445799798705, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.03381108399480581, "state_encode_seconds": 0.06892533300560899, "query_embed_seconds": 0.0046902920003049076, "trained": false, "head_seconds": 0.0012825830199290067, "total_seconds": 0.10873604199150577, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.034481125010643154, "state_encode_seconds": 0.07596966600976884, "query_embed_seconds": 0.001452833996154368, "trained": false, "head_seconds": 0.004837042011786252, "total_seconds": 0.11676662499667145, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.031097958999453112, "state_encode_seconds": 0.06786808298784308, "query_embed_seconds": 0.0010462920181453228, "trained": false, "head_seconds": 0.001347583020105958, "total_seconds": 0.10138150001876056, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.03397670900449157, "state_encode_seconds": 0.06889787499676459, "query_embed_seconds": 0.004996582982130349, "trained": false, "head_seconds": 0.0013357079878915101, "total_seconds": 0.10923783300677314, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.03397670900449157, "state_encode_seconds": 0.06892533300560899, "query_embed_seconds": 0.001717915991321206, "head_seconds": 0.001347583020105958, "total_seconds": 0.10923783300677314 }, "max_seconds": 0.13189445799798705, "under_300ms": 5 }, { "state": "speech", "fields": 28, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.04044641598011367, "state_encode_seconds": 0.0759567920176778, "query_embed_seconds": 0.0018596249865368009, "trained": false, "head_seconds": 0.0034918750170618296, "total_seconds": 0.12178604197106324, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.03919237499940209, "state_encode_seconds": 0.07288812499609776, "query_embed_seconds": 0.0014115410158410668, "trained": false, "head_seconds": 0.005408208002336323, "total_seconds": 0.11891929100966081, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.03518104201066308, "state_encode_seconds": 0.07236087499768473, "query_embed_seconds": 0.00147758299135603, "trained": false, "head_seconds": 0.0015255420003086329, "total_seconds": 0.11056641701725312, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.0331316250085365, "state_encode_seconds": 0.06625862498185597, "query_embed_seconds": 0.0013930420100223273, "trained": false, "head_seconds": 0.00567579100606963, "total_seconds": 0.10647445800714195, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 86, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.034403916011797264, "state_encode_seconds": 0.0746101249824278, "query_embed_seconds": 0.0016509590204805136, "trained": false, "head_seconds": 0.005872208013897762, "total_seconds": 0.11655445900396444, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.03518104201066308, "state_encode_seconds": 0.07288812499609776, "query_embed_seconds": 0.00147758299135603, "head_seconds": 0.005408208002336323, "total_seconds": 0.11655445900396444 }, "max_seconds": 0.12178604197106324, "under_300ms": 5 }, { "state": "video", "fields": 1, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.04310149999218993, "state_encode_seconds": 0.3550283340155147, "query_embed_seconds": 0.0004814579733647406, "trained": false, "head_seconds": 0.0010193330235779285, "total_seconds": 0.39967408298980445, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.04277570900740102, "state_encode_seconds": 0.35542016601539217, "query_embed_seconds": 0.0004606249858625233, "trained": false, "head_seconds": 0.0010592499747872353, "total_seconds": 0.39973408298101276, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.04345179200754501, "state_encode_seconds": 0.347492874978343, "query_embed_seconds": 0.00042116601252928376, "trained": false, "head_seconds": 0.0010006660013459623, "total_seconds": 0.39238404101342894, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.04379708299529739, "state_encode_seconds": 0.34282120800344273, "query_embed_seconds": 0.00041491701267659664, "trained": false, "head_seconds": 0.0009915000118780881, "total_seconds": 0.38804325001547113, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.0442804159829393, "state_encode_seconds": 0.38027550000697374, "query_embed_seconds": 0.00045662501361221075, "trained": false, "head_seconds": 0.0011488330201245844, "total_seconds": 0.4261799579835497, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.04345179200754501, "state_encode_seconds": 0.3550283340155147, "query_embed_seconds": 0.00045662501361221075, "head_seconds": 0.0010193330235779285, "total_seconds": 0.39967408298980445 }, "max_seconds": 0.4261799579835497, "under_300ms": 0 }, { "state": "video", "fields": 4, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.04463912500068545, "state_encode_seconds": 0.3781175419862848, "query_embed_seconds": 0.0006854170060250908, "trained": false, "head_seconds": 0.0011744580115191638, "total_seconds": 0.4246371249901131, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.04435204100445844, "state_encode_seconds": 0.3557324589928612, "query_embed_seconds": 0.0005507909809239209, "trained": false, "head_seconds": 0.001133959012804553, "total_seconds": 0.40179754100972787, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.042686999979196116, "state_encode_seconds": 0.34845437499461696, "query_embed_seconds": 0.0006089580128900707, "trained": false, "head_seconds": 0.0024466670001856983, "total_seconds": 0.3942195840063505, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.04590104197268374, "state_encode_seconds": 0.39702925001620315, "query_embed_seconds": 0.0006155410083010793, "trained": false, "head_seconds": 0.0017480840033385903, "total_seconds": 0.4453298329899553, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.042907666007522494, "state_encode_seconds": 0.34891875000903383, "query_embed_seconds": 0.0005797499907203019, "trained": false, "head_seconds": 0.004959250014508143, "total_seconds": 0.3973861249978654, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.04435204100445844, "state_encode_seconds": 0.3557324589928612, "query_embed_seconds": 0.0006089580128900707, "head_seconds": 0.0017480840033385903, "total_seconds": 0.40179754100972787 }, "max_seconds": 0.4453298329899553, "under_300ms": 0 }, { "state": "video", "fields": 16, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.04560420798952691, "state_encode_seconds": 0.34458412500680424, "query_embed_seconds": 0.0011750420089811087, "trained": false, "head_seconds": 0.005385875003412366, "total_seconds": 0.3967709169955924, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.0461367919924669, "state_encode_seconds": 0.3806338330032304, "query_embed_seconds": 0.0010893750004470348, "trained": false, "head_seconds": 0.005805917026009411, "total_seconds": 0.43368679197737947, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.04518658298184164, "state_encode_seconds": 0.4042663339932915, "query_embed_seconds": 0.001439208019291982, "trained": false, "head_seconds": 0.0027748749998863786, "total_seconds": 0.45374433300457895, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.0466007910144981, "state_encode_seconds": 0.36523454199777916, "query_embed_seconds": 0.0015491669764742255, "trained": false, "head_seconds": 0.001510500005679205, "total_seconds": 0.41491662501357496, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.04235237502143718, "state_encode_seconds": 0.39170579198980704, "query_embed_seconds": 0.0011479579843580723, "trained": false, "head_seconds": 0.0015190000121947378, "total_seconds": 0.4367624159785919, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.04560420798952691, "state_encode_seconds": 0.3806338330032304, "query_embed_seconds": 0.0011750420089811087, "head_seconds": 0.0027748749998863786, "total_seconds": 0.43368679197737947 }, "max_seconds": 0.45374433300457895, "under_300ms": 0 }, { "state": "video", "fields": 28, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.04438375000609085, "state_encode_seconds": 0.436862457980169, "query_embed_seconds": 0.004034208017401397, "trained": false, "head_seconds": 0.00406870900769718, "total_seconds": 0.4894299999868963, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.04700783401494846, "state_encode_seconds": 0.4069262080010958, "query_embed_seconds": 0.0014327079989016056, "trained": false, "head_seconds": 0.00551750001613982, "total_seconds": 0.46090845798607916, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.04538316698744893, "state_encode_seconds": 0.3784948750108015, "query_embed_seconds": 0.001702665991615504, "trained": false, "head_seconds": 0.0020127500174567103, "total_seconds": 0.4276172919780947, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.046234749985160306, "state_encode_seconds": 0.42870437502278946, "query_embed_seconds": 0.004925999994156882, "trained": false, "head_seconds": 0.0024462919973302633, "total_seconds": 0.48236045899102464, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 328, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.04259108399855904, "state_encode_seconds": 0.34766154098906554, "query_embed_seconds": 0.0016559170035179704, "trained": false, "head_seconds": 0.005331209016731009, "total_seconds": 0.3972586249874439, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.04538316698744893, "state_encode_seconds": 0.4069262080010958, "query_embed_seconds": 0.001702665991615504, "head_seconds": 0.00406870900769718, "total_seconds": 0.46090845798607916 }, "max_seconds": 0.4894299999868963, "under_300ms": 0 }, { "state": "video_speech", "fields": 1, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.07553274999372661, "state_encode_seconds": 0.42347375000827014, "query_embed_seconds": 0.00045383398537524045, "trained": false, "head_seconds": 0.0010031670099124312, "total_seconds": 0.5004834170104004, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.07891866701538675, "state_encode_seconds": 0.43575612499262206, "query_embed_seconds": 0.00045900000259280205, "trained": false, "head_seconds": 0.001060832990333438, "total_seconds": 0.5162106659845449, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.07799137500114739, "state_encode_seconds": 0.4307675420132, "query_embed_seconds": 0.0011055829818360507, "trained": false, "head_seconds": 0.0048137079866137356, "total_seconds": 0.5147268339933362, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.07632429199293256, "state_encode_seconds": 0.42008512499160133, "query_embed_seconds": 0.0004938330093864352, "trained": false, "head_seconds": 0.0010908329859375954, "total_seconds": 0.49801129099796526, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 1, "candidate_count": 3, "query_tokens": 59, "padded_query_tokens": 60, "state_memory_batch_size": 1, "preprocess_seconds": 0.07897920798859559, "state_encode_seconds": 0.4273338750062976, "query_embed_seconds": 0.00042916700476780534, "trained": false, "head_seconds": 0.0010513330053072423, "total_seconds": 0.5078184159938246, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.07799137500114739, "state_encode_seconds": 0.4273338750062976, "query_embed_seconds": 0.00045900000259280205, "head_seconds": 0.001060832990333438, "total_seconds": 0.5078184159938246 }, "max_seconds": 0.5162106659845449, "under_300ms": 0 }, { "state": "video_speech", "fields": 4, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.08018525000079535, "state_encode_seconds": 0.41756066601374187, "query_embed_seconds": 0.0005665839998982847, "trained": false, "head_seconds": 0.0011635419796220958, "total_seconds": 0.49949158300296403, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.07973349999519996, "state_encode_seconds": 0.4295728750003036, "query_embed_seconds": 0.0006390829803422093, "trained": false, "head_seconds": 0.0012088750081602484, "total_seconds": 0.5111723330046516, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.09626458398997784, "state_encode_seconds": 0.4280259160150308, "query_embed_seconds": 0.0005661250033881515, "trained": false, "head_seconds": 0.001206041983095929, "total_seconds": 0.5260827919992153, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.0781442919978872, "state_encode_seconds": 0.4153163750015665, "query_embed_seconds": 0.0005662920011673123, "trained": false, "head_seconds": 0.0028537499892991036, "total_seconds": 0.4969030420179479, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 4, "candidate_count": 10, "query_tokens": 184, "padded_query_tokens": 200, "state_memory_batch_size": 1, "preprocess_seconds": 0.07937216700520366, "state_encode_seconds": 0.41883766700630076, "query_embed_seconds": 0.0006210409919731319, "trained": false, "head_seconds": 0.0012313330080360174, "total_seconds": 0.5000778750109021, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.07973349999519996, "state_encode_seconds": 0.41883766700630076, "query_embed_seconds": 0.0005665839998982847, "head_seconds": 0.0012088750081602484, "total_seconds": 0.5000778750109021 }, "max_seconds": 0.5260827919992153, "under_300ms": 0 }, { "state": "video_speech", "fields": 16, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.07921012499718927, "state_encode_seconds": 0.42091008397983387, "query_embed_seconds": 0.0012018750130664557, "trained": false, "head_seconds": 0.005473249999340624, "total_seconds": 0.5068114580062684, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.07781325001269579, "state_encode_seconds": 0.42197975001181476, "query_embed_seconds": 0.0010271249921061099, "trained": false, "head_seconds": 0.005172542005311698, "total_seconds": 0.5060163329762872, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.07651900002383627, "state_encode_seconds": 0.416757457976928, "query_embed_seconds": 0.004449000000022352, "trained": false, "head_seconds": 0.0015247499977704138, "total_seconds": 0.49926879099803045, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.07601066699135117, "state_encode_seconds": 0.41538487499929033, "query_embed_seconds": 0.0010292080114595592, "trained": false, "head_seconds": 0.005016874987632036, "total_seconds": 0.4974667080095969, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 16, "candidate_count": 40, "query_tokens": 736, "padded_query_tokens": 800, "state_memory_batch_size": 1, "preprocess_seconds": 0.07797779198153876, "state_encode_seconds": 0.4197924170002807, "query_embed_seconds": 0.0010504999954719096, "trained": false, "head_seconds": 0.0015509170189034194, "total_seconds": 0.5003884169855155, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.07781325001269579, "state_encode_seconds": 0.4197924170002807, "query_embed_seconds": 0.0010504999954719096, "head_seconds": 0.005016874987632036, "total_seconds": 0.5003884169855155 }, "max_seconds": 0.5068114580062684, "under_300ms": 0 }, { "state": "video_speech", "fields": 28, "attempted": 5, "completed": 5, "samples": [ { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.07941245898837224, "state_encode_seconds": 0.423694750003051, "query_embed_seconds": 0.004491458006668836, "trained": false, "head_seconds": 0.0019519590132404119, "total_seconds": 0.5095802079886198, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.07987937500001863, "state_encode_seconds": 0.4265894999844022, "query_embed_seconds": 0.0013566250272560865, "trained": false, "head_seconds": 0.001676957996096462, "total_seconds": 0.5095168749976438, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.07661350001581013, "state_encode_seconds": 0.4164905420038849, "query_embed_seconds": 0.0051647499785758555, "trained": false, "head_seconds": 0.002027959009865299, "total_seconds": 0.5003168340190314, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.07714025001041591, "state_encode_seconds": 0.42116833399632014, "query_embed_seconds": 0.0013382909819483757, "trained": false, "head_seconds": 0.00171387501177378, "total_seconds": 0.5013800839951728, "status": "runtime_probe_only" }, { "execution": "shared_state_parallel_decision_head", "state_encodes": 1, "state_tokens": 431, "state_layers": 4, "compute_dtype": "float16", "field_count": 28, "candidate_count": 70, "query_tokens": 1288, "padded_query_tokens": 1400, "state_memory_batch_size": 1, "preprocess_seconds": 0.07797850001952611, "state_encode_seconds": 0.42002383299404755, "query_embed_seconds": 0.0014834169996902347, "trained": false, "head_seconds": 0.005803791980724782, "total_seconds": 0.5053050830028951, "status": "runtime_probe_only" } ], "median_seconds": { "preprocess_seconds": 0.07797850001952611, "state_encode_seconds": 0.42116833399632014, "query_embed_seconds": 0.0014834169996902347, "head_seconds": 0.0019519590132404119, "total_seconds": 0.5053050830028951 }, "max_seconds": 0.5095802079886198, "under_300ms": 0 } ] }