alvarobartt's picture
download
raw
13.3 kB
{
"schema_version": 1,
"null_semantics": "Unavailable, not zero. Empty CSV cells represent null.",
"rounding": "Native GiB/GB logs generally have 0.01 GiB resolution. NVML snapshots have 1 MiB resolution. llama.cpp buffer logs generally have 0.01 MiB resolution. Conversion to integer bytes does not increase measurement precision.",
"unit_conversion": {
"GiB": 1073741824,
"MiB": 1048576,
"SGLang_native_GB": 1073741824
},
"measurement_scope": "All primary measurements are device 0 on a single GPU. Estimates use batch 1 at requested context. Historical measurements and newly pinned model revisions need not be identical.",
"columns": {
"activation_peak_reported_bytes": {
"description": "vLLM profiler activation peak; not an independent ready-state allocation.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"arguments": {
"description": "Archived engine CLI arguments for this configuration.",
"unit": null,
"nullable": true,
"example_type": "list"
},
"batch_size": {
"description": "Batch size.",
"unit": null,
"nullable": true,
"example_type": "int"
},
"chunked_prefill_tokens": {
"description": "SGLang native chunked-prefill budget; used as approximate hf-mem token budget.",
"unit": null,
"nullable": true,
"example_type": "int"
},
"compute_buffer_reported_bytes": {
"description": "llama.cpp main decoder compute buffer; not added to residual.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"config": {
"description": "Named native launch mode. text-only disables native multimodal encoders.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"cuda_graph_decode_reported_bytes": {
"description": "SGLang decode graph capture memory delta, rounded.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"cuda_graph_engine_estimate_bytes": {
"description": "vLLM graph pool estimate, not the actual measured allocation.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"cuda_graph_prefill_reported_bytes": {
"description": "SGLang prefill graph capture memory delta, rounded.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"cuda_graph_reported_bytes": {
"description": "vLLM actual graph pool allocation reported after capture, rounded.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"device_capacity_bytes": {
"description": "Total GPU memory reported by nvidia-smi, not necessarily CUDA-allocatable capacity.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"device_index": {
"description": "Device index.",
"unit": null,
"nullable": true,
"example_type": "int"
},
"device_probe_bytes": {
"description": "Whole-device memory snapshot after an image/audio probe; may follow a failed probe.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"device_ready_bytes": {
"description": "Whole-device nvidia-smi memory.used snapshot after ready; not process-exclusive and not a peak.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"device_text_bytes": {
"description": "Whole-device memory snapshot after a short text probe; not a peak.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"encoder_compute_buffer_reported_bytes": {
"description": "llama.cpp multimodal encoder compute buffer; not added to residual.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"engine": {
"description": "Engine.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"engine_image": {
"description": "Engine image.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"engine_image_digest": {
"description": "Engine image digest.",
"unit": null,
"nullable": true,
"example_type": null
},
"engine_version": {
"description": "Native logged engine version or build; null if not recorded.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"estimate_error": {
"description": "Estimate error.",
"unit": null,
"nullable": true,
"example_type": null
},
"estimation_assumptions": {
"description": "Run-specific missing information, scope differences and token-budget approximations.",
"unit": null,
"nullable": true,
"example_type": null
},
"evidence": {
"description": "Relative path to selected sanitized native log lines.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"file": {
"description": "File.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"gpu_memory_utilization": {
"description": "Gpu memory utilization.",
"unit": null,
"nullable": true,
"example_type": "float"
},
"hardware": {
"description": "Hardware.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"hardware_flavor": {
"description": "Hardware flavor.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"healthy": {
"description": "Whether the archived run reached a ready/healthy server; independent of inference probe success.",
"unit": null,
"nullable": true,
"example_type": "bool"
},
"hf_mem_model_revision": {
"description": "Immutable model commit SHA used for the newly generated estimate.",
"unit": null,
"nullable": true,
"example_type": null
},
"hf_mem_offloaded_bytes": {
"description": "hf-mem separately reported offloaded tensors; not blindly added to or subtracted from weights.",
"unit": "bytes",
"nullable": true,
"example_type": null
},
"hf_mem_overhead_activations_bytes": {
"description": "Overhead activations bytes. Native hf-mem overhead component.",
"unit": "bytes",
"nullable": true,
"example_type": null
},
"hf_mem_overhead_bytes": {
"description": "hf-mem estimated overhead including applicable activation, encoder, graph and runtime allowances.",
"unit": "bytes",
"nullable": true,
"example_type": null
},
"hf_mem_overhead_encoder_bytes": {
"description": "Overhead encoder bytes. Native hf-mem overhead component.",
"unit": "bytes",
"nullable": true,
"example_type": null
},
"hf_mem_overhead_graphs_bytes": {
"description": "Overhead graphs bytes. Native hf-mem overhead component.",
"unit": "bytes",
"nullable": true,
"example_type": null
},
"hf_mem_overhead_runtime_bytes": {
"description": "Overhead runtime bytes. Native hf-mem overhead component.",
"unit": "bytes",
"nullable": true,
"example_type": null
},
"hf_mem_requested_batch_size": {
"description": "Requested batch size.",
"unit": null,
"nullable": true,
"example_type": null
},
"hf_mem_requested_kv_bytes": {
"description": "Estimated cache for batch size 1 at the requested maximum context, not the runtime full cache pool.",
"unit": "bytes",
"nullable": true,
"example_type": null
},
"hf_mem_requested_total_bytes": {
"description": "hf-mem total for its own request-context estimation scope; not directly comparable to a full cache-pool snapshot.",
"unit": "bytes",
"nullable": true,
"example_type": null
},
"hf_mem_version": {
"description": "Version.",
"unit": null,
"nullable": true,
"example_type": null
},
"hf_mem_weights_bytes": {
"description": "Weights as reported by hf-mem; includes model components such as multimodal projectors according to its format rules.",
"unit": "bytes",
"nullable": true,
"example_type": null
},
"kv_cache_dtype": {
"description": "Kv cache dtype.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"kv_pool_reported_bytes": {
"description": "Full allocated/reported native KV cache pool, not one request. llama.cpp attention buffers summed after real loading.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"kv_pool_tokens": {
"description": "Native reported cache capacity in tokens; hybrid attention groups may not map linearly to request memory.",
"unit": null,
"nullable": true,
"example_type": "int"
},
"max_model_len": {
"description": "Requested maximum context length in tokens.",
"unit": null,
"nullable": true,
"example_type": "int"
},
"max_num_batched_tokens": {
"description": "vLLM native batched token limit if logged; null does not imply the hf-mem default matched.",
"unit": null,
"nullable": true,
"example_type": "int"
},
"max_num_seqs": {
"description": "Max num seqs.",
"unit": null,
"nullable": true,
"example_type": null
},
"measurement_date": {
"description": "Measurement date.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"measurement_date_basis": {
"description": "Measurement date basis.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"mem_fraction_static": {
"description": "Mem fraction static.",
"unit": null,
"nullable": true,
"example_type": "float"
},
"model": {
"description": "Model.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"model_revision": {
"description": "Historical model commit SHA when recoverable from the native log; otherwise null.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"model_revision_status": {
"description": "Model revision status.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"multimodal_probe": {
"description": "Image/audio probe outcome as image:ok, image:fail, audio:ok or audio:fail; null means no probe recorded.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"non_torch_derived_bytes": {
"description": "vLLM consumed memory minus model-loading memory; derived from rounded log values.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"other_device_memory_residual_bytes": {
"description": "Derived ready snapshot minus reported weights, cache pool, and separately reported recurrent state. Accounting remainder, not pure overhead.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"recurrent_state_reported_bytes": {
"description": "llama.cpp native CUDA recurrent state buffers; separate from attention KV bytes.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"run_id": {
"description": "Unique archived launch and configuration identifier; repeated model IDs are expected.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"source_file": {
"description": "Source file.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"source_lines": {
"description": "Number of lines in the private source log; evidence line prefixes refer to those line numbers.",
"unit": null,
"nullable": true,
"example_type": "int"
},
"source_sha256": {
"description": "SHA-256 of private complete native source log, before excerpt selection and sanitization.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"tensor_parallel_size": {
"description": "Tensor parallel size.",
"unit": null,
"nullable": true,
"example_type": "int"
},
"text_request_succeeded": {
"description": "Short text probe completion status; null means no such probe was recorded.",
"unit": null,
"nullable": true,
"example_type": "bool"
},
"weights_dtype": {
"description": "Weights dtype.",
"unit": null,
"nullable": true,
"example_type": "str"
},
"weights_host_mapped_bytes": {
"description": "llama.cpp CPU-mapped model buffer; not measured resident RAM.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"weights_plus_non_torch_reported_bytes": {
"description": "vLLM consumed memory from native worker summary, rounded.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
},
"weights_reported_bytes": {
"description": "Engine loading memory delta or llama.cpp CUDA model buffer; rounded and engine-specific.",
"unit": "bytes",
"nullable": true,
"example_type": "int"
}
}
}

Xet Storage Details

Size:
13.3 kB
·
Xet hash:
e021a8c87c8b352f9881b63a026b8b8daf502699c9d6233e30756d0440f9dca8

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.