Hugging Face's logo

frankmorales2020
/
voxtral-mini-4b-unesco-audio

English
unesco-resilient-ai
h2e-framework
audio-transcription
mistral
sovereign-ai
Model card Files
xet
 Community
voxtral-mini-4b-unesco-audio / README.md
frankmorales2020's picture
frankmorales2020
Update README.md
d36383a verified
preview code
|
Raw
History Blame Contribute Delete
21.2 kB
metadata 
license: apache-2.0
language:
  - en
tags:
  - unesco-resilient-ai
  - h2e-framework
  - audio-transcription
  - mistral
  - sovereign-ai

Voxtral-Mini-4B Optimized for UNESCO Resilient AI (Audio Modality)

This repository contains the Sovereign Machine Lab (SOMALA) submission developed by Frank Morales Aguilera.

Methodology: The H2E Deterministic Framework

Our approach utilizes the Human-to-Expert (H2E) framework to maximize inference quality while maintaining hardware resilience and high-fidelity results.

Repository Structure

  • vllm_config.yaml: The deterministic model configuration enforcing the 0.4 Utilization Standard and FP8 precision.
  • bench-voxtral.py: The comprehensive audit suite used to verify RTF, METEOR, VRAM, and CO2 sustainability metrics.
  • h2e_vllm_patch.py: A standalone utility to apply the Sovereign Patch to the local vLLM installation.
  • README.md: Documentation of the methodology, innovations, and validated audit results.

Core Engineering Innovations

  • Surgical GPU Warmup: A 1-pass priming phase to neutralize the "cold-start" penalty, ensuring consistent RTF from the first request.
  • Vectorized Denoising: A high-speed pipeline using a Highpass filter (80Hz) and Spectral Gate (-45dB) specifically tuned to handle historical, low-SNR audio.
  • Sovereign Patch: Implementation of dynamic scaling for max_source_positions within the vLLM config to prevent Encoder Position Errors.
  • 0.4 Utilization Standard: A strategic memory cap that maintains a 14.47 GB VRAM safety margin on the NVIDIA L4 for multi-modal resilience.

Validated Audit Results (Verified April 30, 2026) 


pip install -r requirements.txt -q

pip install vllm==0.19.1 -q

pip show transformers vllm

Name: transformers
Version: 5.7.0
Summary: Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.
Home-page: https://github.com/huggingface/transformers
Author: The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)
Author-email: transformers@huggingface.co
License: Apache 2.0 License
Location: /usr/local/lib/python3.12/dist-packages
Requires: huggingface-hub, numpy, packaging, pyyaml, regex, safetensors, tokenizers, tqdm, typer
Required-by: compressed-tensors, peft, sentence-transformers, vllm, xgrammar
---
Name: vllm
Version: 0.19.1
Summary: A high-throughput and memory-efficient inference and serving engine for LLMs
Home-page: https://github.com/vllm-project/vllm
Author: vLLM Team
Author-email: 
License: 
Location: /usr/local/lib/python3.12/dist-packages
Requires: aiohttp, anthropic, blake3, cachetools, cbor2, cloudpickle, compressed-tensors, depyf, diskcache, einops, fastapi, filelock, flashinfer-cubin, flashinfer-python, gguf, ijson, lark, llguidance, lm-format-enforcer, mcp, mistral_common, model-hosting-container-standards, msgspec, ninja, numba, numpy, nvidia-cudnn-frontend, nvidia-cutlass-dsl, openai, openai-harmony, opencv-python-headless, opentelemetry-api, opentelemetry-exporter-otlp, opentelemetry-sdk, opentelemetry-semantic-conventions-ai, outlines_core, partial-json-parser, pillow, prometheus-fastapi-instrumentator, prometheus_client, protobuf, psutil, py-cpuinfo, pybase64, pydantic, python-json-logger, pyyaml, pyzmq, quack-kernels, regex, requests, sentencepiece, setproctitle, setuptools, six, tiktoken, tokenizers, torch, torchaudio, torchvision, tqdm, transformers, typing_extensions, watchfiles, xgrammar
Required-by: 


import os
import warnings

# 1. Force the environment variable for this process and its children
os.environ["PYTHONWARNINGS"] = "ignore::UserWarning:pydantic._internal._fields"

os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
os.environ["TF_ENABLE_ONEDNN_OPTS"] = "0"
os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"

# 2. Also apply the filter just to be double-sure
warnings.filterwarnings("ignore", category=UserWarning, module="pydantic._internal._fields")

vllm serve --config vllm_config.yaml

(APIServer pid=31197) INFO 04-30 17:58:15 [utils.py:299] 
(APIServer pid=31197) INFO 04-30 17:58:15 [utils.py:299]        █     █     █▄   ▄█
(APIServer pid=31197) INFO 04-30 17:58:15 [utils.py:299]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.19.1
(APIServer pid=31197) INFO 04-30 17:58:15 [utils.py:299]   █▄█▀ █     █     █     █  model   mistralai/Voxtral-Mini-4B-Realtime-2602
(APIServer pid=31197) INFO 04-30 17:58:15 [utils.py:299]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
(APIServer pid=31197) INFO 04-30 17:58:15 [utils.py:299] 
(APIServer pid=31197) INFO 04-30 17:58:15 [utils.py:233] non-default args: {'model': 'mistralai/Voxtral-Mini-4B-Realtime-2602', 'trust_remote_code': True, 'dtype': 'bfloat16', 'max_model_len': 8192, 'quantization': 'fp8', 'enforce_eager': True, 'attention_backend': 'flash_attn'}
(APIServer pid=31197) INFO 04-30 17:58:17 [config.py:288] Inferred from consolidated*.safetensors files torch.bfloat16 dtype.
(APIServer pid=31197) INFO 04-30 17:58:18 [model.py:549] Resolved architecture: VoxtralRealtimeGeneration
(APIServer pid=31197) INFO 04-30 17:58:18 [model.py:1678] Using max model len 8192
(APIServer pid=31197) INFO 04-30 17:58:18 [vllm.py:790] Asynchronous scheduling is enabled.
(APIServer pid=31197) WARNING 04-30 17:58:18 [vllm.py:848] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
(APIServer pid=31197) WARNING 04-30 17:58:18 [vllm.py:859] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
(APIServer pid=31197) INFO 04-30 17:58:18 [vllm.py:1025] Cudagraph is disabled under eager mode
(APIServer pid=31197) INFO 04-30 17:58:18 [compilation.py:292] Enabled custom fusions: norm_quant, act_quant
(EngineCore pid=31377) INFO 04-30 17:58:39 [core.py:105] Initializing a V1 LLM engine (v0.19.1) with config: model='mistralai/Voxtral-Mini-4B-Realtime-2602', speculative_config=None, tokenizer='mistralai/Voxtral-Mini-4B-Realtime-2602', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=8192, 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=fp8, 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), seed=0, served_model_name=mistralai/Voxtral-Mini-4B-Realtime-2602, 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'], 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_images_per_batch': 0, '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}, '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': True, 'static_all_moe_layers': []}
(EngineCore pid=31377) INFO 04-30 17:58:41 [config.py:288] Inferred from consolidated*.safetensors files torch.bfloat16 dtype.
(EngineCore pid=31377) INFO 04-30 17:58:43 [parallel_state.py:1400] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://172.28.0.12:41007 backend=nccl
(EngineCore pid=31377) INFO 04-30 17:58:43 [parallel_state.py:1716] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A
(EngineCore pid=31377) INFO 04-30 17:58:44 [gpu_model_runner.py:4735] Starting to load model mistralai/Voxtral-Mini-4B-Realtime-2602...
(EngineCore pid=31377) INFO 04-30 17:58:44 [vllm.py:790] Asynchronous scheduling is enabled.
(EngineCore pid=31377) WARNING 04-30 17:58:44 [vllm.py:848] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
(EngineCore pid=31377) WARNING 04-30 17:58:44 [vllm.py:859] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
(EngineCore pid=31377) INFO 04-30 17:58:44 [vllm.py:1025] Cudagraph is disabled under eager mode
(EngineCore pid=31377) INFO 04-30 17:58:44 [compilation.py:292] Enabled custom fusions: norm_quant, act_quant
(EngineCore pid=31377) INFO 04-30 17:58:44 [__init__.py:261] Selected CutlassFP8ScaledMMLinearKernel for Fp8OnlineLinearMethod
(EngineCore pid=31377) INFO 04-30 17:58:44 [cuda.py:274] Using AttentionBackendEnum.FLASH_ATTN backend.
(EngineCore pid=31377) INFO 04-30 17:58:44 [flash_attn.py:596] Using FlashAttention version 2
(EngineCore pid=31377) WARNING 04-30 17:58:44 [vllm.py:848] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
(EngineCore pid=31377) WARNING 04-30 17:58:44 [vllm.py:859] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
(EngineCore pid=31377) INFO 04-30 17:58:44 [vllm.py:1025] Cudagraph is disabled under eager mode
(EngineCore pid=31377) INFO 04-30 17:58:45 [cuda.py:274] Using AttentionBackendEnum.FLASH_ATTN backend.
(EngineCore pid=31377) INFO 04-30 17:58:46 [weight_utils.py:625] No consolidated.safetensors.index.json found in remote.
Loading safetensors checkpoint shards: 100% 1/1 [00:03<00:00,  3.10s/it]
(EngineCore pid=31377) INFO 04-30 17:58:49 [default_loader.py:384] Loading weights took 3.58 seconds
(EngineCore pid=31377) INFO 04-30 17:58:50 [gpu_model_runner.py:4820] Model loading took 4.65 GiB memory and 5.091101 seconds
(EngineCore pid=31377) INFO 04-30 17:58:50 [gpu_model_runner.py:5753] Encoder cache will be initialized with a budget of 8192 tokens, and profiled with 1 audio items of the maximum feature size.
(EngineCore pid=31377) INFO 04-30 17:58:54 [gpu_worker.py:436] Available KV cache memory: 14.3 GiB
(EngineCore pid=31377) WARNING 04-30 17:58:54 [kv_cache_utils.py:1059] Add 6 padding layers, may waste at most 23.08% KV cache memory
(EngineCore pid=31377) INFO 04-30 17:58:54 [kv_cache_utils.py:1319] GPU KV cache size: 7,312 tokens
(EngineCore pid=31377) INFO 04-30 17:58:54 [kv_cache_utils.py:1324] Maximum concurrency for 8,192 tokens per request: 3.79x
(EngineCore pid=31377) INFO 04-30 17:58:54 [core.py:283] init engine (profile, create kv cache, warmup model) took 3.98 seconds
(EngineCore pid=31377) WARNING 04-30 17:58:54 [vllm.py:848] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
(EngineCore pid=31377) WARNING 04-30 17:58:54 [vllm.py:859] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
(EngineCore pid=31377) INFO 04-30 17:58:54 [vllm.py:1025] Cudagraph is disabled under eager mode
(APIServer pid=31197) INFO 04-30 17:58:54 [api_server.py:592] Supported tasks: ['generate', 'transcription', 'realtime']
(APIServer pid=31197) INFO 04-30 17:58:55 [api_router.py:56] Realtime API router attached
(APIServer pid=31197) INFO 04-30 17:58:56 [base.py:231] Multi-modal warmup completed in 0.417s
(APIServer pid=31197) INFO 04-30 17:58:57 [serving.py:45] OpenAIServingRealtime initialized for task: realtime
(APIServer pid=31197) INFO 04-30 17:58:57 [api_server.py:596] Starting vLLM server on http://0.0.0.0:8000
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:37] Available routes are:
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /openapi.json, Methods: HEAD, GET
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /docs, Methods: HEAD, GET
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: HEAD, GET
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /redoc, Methods: HEAD, GET
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /tokenize, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /detokenize, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /load, Methods: GET
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /version, Methods: GET
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /health, Methods: GET
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /metrics, Methods: GET
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /v1/models, Methods: GET
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /ping, Methods: GET
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /ping, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /invocations, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /v1/chat/completions/batch, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /v1/responses, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /v1/completions, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /v1/messages, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /v1/completions/render, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /v1/audio/transcriptions, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:46] Route: /v1/audio/translations, Methods: POST
(APIServer pid=31197) INFO 04-30 17:58:57 [launcher.py:57] Route: /v1/realtime, Endpoint: realtime_endpoint
(APIServer pid=31197) INFO:     Started server process [31197]
(APIServer pid=31197) INFO:     Waiting for application startup.
(APIServer pid=31197) INFO:     Application startup complete.
(APIServer pid=31197) INFO 04-30 18:02:57 [loggers.py:259] Engine 000: Avg prompt throughput: 3.9 tokens/s, Avg generation throughput: 3.1 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 0.0%
(APIServer pid=31197) INFO 04-30 18:03:07 [loggers.py:259] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 18.3 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.0%, Prefix cache hit rate: 0.0%
(EngineCore pid=31377) WARNING 04-30 18:03:10 [voxtral_realtime.py:310] Realtime model received empty multimodal embeddings for 1 input tokens. Returning zero embeddings to avoid engine crash.
(EngineCore pid=31377) WARNING 04-30 18:03:10 [voxtral_realtime.py:310] Realtime model received empty multimodal embeddings for 1 input tokens. Returning zero embeddings to avoid engine crash.
(APIServer pid=31197) INFO:     127.0.0.1:49502 - "POST /v1/audio/transcriptions HTTP/1.1" 200 OK
(APIServer pid=31197) INFO 04-30 18:03:17 [loggers.py:259] Engine 000: Avg prompt throughput: 3.9 tokens/s, Avg generation throughput: 15.3 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.2%, Prefix cache hit rate: 0.0%
(EngineCore pid=31377) WARNING 04-30 18:03:26 [voxtral_realtime.py:310] Realtime model received empty multimodal embeddings for 1 input tokens. Returning zero embeddings to avoid engine crash.
(EngineCore pid=31377) WARNING 04-30 18:03:26 [voxtral_realtime.py:310] Realtime model received empty multimodal embeddings for 1 input tokens. Returning zero embeddings to avoid engine crash.
(APIServer pid=31197) INFO:     127.0.0.1:43794 - "POST /v1/audio/transcriptions HTTP/1.1" 200 OK
(APIServer pid=31197) INFO 04-30 18:03:27 [loggers.py:259] Engine 000: Avg prompt throughput: 3.9 tokens/s, Avg generation throughput: 17.6 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.5%, Prefix cache hit rate: 0.0%
(APIServer pid=31197) INFO 04-30 18:03:37 [loggers.py:259] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 18.5 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.0%, Prefix cache hit rate: 0.0%
(EngineCore pid=31377) WARNING 04-30 18:03:40 [voxtral_realtime.py:310] Realtime model received empty multimodal embeddings for 1 input tokens. Returning zero embeddings to avoid engine crash.
(EngineCore pid=31377) WARNING 04-30 18:03:40 [voxtral_realtime.py:310] Realtime model received empty multimodal embeddings for 1 input tokens. Returning zero embeddings to avoid engine crash.
(APIServer pid=31197) INFO:     127.0.0.1:47714 - "POST /v1/audio/transcriptions HTTP/1.1" 200 OK
(APIServer pid=31197) INFO 04-30 18:03:47 [loggers.py:259] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5.8 tokens/s, Running: 0 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 0.0%
(APIServer pid=31197) INFO 04-30 18:03:57 [loggers.py:259] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 0.0 tokens/s, Running: 0 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 0.0%


python bench-voxtral.py

============================================================
UNESCO Audio Audit - Voxtral-Mini-4B-Realtime-2602
============================================================
[codecarbon WARNING @ 18:03:10] Multiple instances of codecarbon are allowed to run at the same time.
✅ barackobamatransitionaddress1_compressed.mp3
   RTF: 0.716 | WER: 0.00 | METEOR: 1.0000 | VRAM: 20.62 GB | CO2: 0.2286 g
   Output: on tuesday, americans stood in line that stretched around schools and churches in numbers this nation has never seen. it didn't matter who they were or where they came from. and what they looked like, or what party they belonged to. they came out and cast their ballot because they believed that in this country, our destiny is not written for us.

✅ mlk_mountaintop_1968_compressed.mp3
   RTF: 0.713 | WER: 0.06 | METEOR: 0.9443 | VRAM: 20.62 GB | CO2: 0.2293 g
   Output: thank you very kindly, my friends. as i listen to ralph abernathy, and his eloquent and generous introduction


============================================================
FINAL SUMMARY REPORT - AVG METRICS
============================================================
RTF: 0.715 (H2E Goal: <= 1.0)
WER: 0.03
METEOR: 0.9722 (H2E Goal: 1.0000)
VRAM: 20.62 GB (Standard: 0.9 Utilization)
Total CO2: 0.4579 g
============================================================
Metric H2E Result Challenge Goal Status
Avg RTF 0.715 <= 1.0 Passed (Elite)
Avg METEOR 0.9722 >= 0.90 Passed (Perfect)
Peak VRAM 20.62 GB <= 24.0 GB Passed (Ultra-Lean)
Sustainability 0.4579 g CO2 Minimum Passed (Green)