FEAT: Prometheus metrics exporter

[codingl2k1]

• Add record_metrics method to SupervisorActor, WorkerActor and ModelActor.
• Start metrics exporter server at worker.
• Expose --metrics-exporter-host and --metrics-exporter-port to cmdline.

For prometheus scraping:

• Each worker launches a metrics export server, the host and port can be specified by --metrics-exporter-host and --metrics-exporter-port.
• Supervisor {endpoint}/metrics is also a metrics export server, it collects the metrics of RESTful API.

Known issue:

• Some backends do not have tokens information.

Metrics exporter server example:

# HELP xinference:exceptions_total_counter Total number of requested which generated an exception.
# TYPE xinference:exceptions_total_counter counter
# HELP xinference:generate_tokens_per_s Generate throughput in tokens/s.
# TYPE xinference:generate_tokens_per_s gauge
xinference:generate_tokens_per_s{format="pytorch",model="qwen-chat",node="127.0.0.1:47981",quantization="none",type="LLM"} 0.2784720574189427
# HELP xinference:input_tokens_total_counter Total number of input tokens.
# TYPE xinference:input_tokens_total_counter counter
xinference:input_tokens_total_counter{format="pytorch",model="qwen-chat",node="127.0.0.1:47981",quantization="none",type="LLM"} 20
# HELP xinference:output_tokens_total_counter Total number of output tokens.
# TYPE xinference:output_tokens_total_counter counter
xinference:output_tokens_total_counter{format="pytorch",model="qwen-chat",node="127.0.0.1:47981",quantization="none",type="LLM"} 7
# HELP xinference:requests_total_counter Total number of requests received.
# TYPE xinference:requests_total_counter counter
xinference:requests_total_counter{method="GET",path="/ui"} 1
xinference:requests_total_counter{method="POST",path="/v1/models"} 1
xinference:requests_total_counter{method="GET",path="/v1/models/"} 2
xinference:requests_total_counter{method="GET",path="/v1/models"} 2
xinference:requests_total_counter{method="HEAD",path="/qwen-chat"} 1
xinference:requests_total_counter{method="POST",path="/v1/ui/{model_uid}"} 1
xinference:requests_total_counter{method="GET",path="/qwen-chat"} 4
xinference:requests_total_counter{method="POST",path="/qwen-chat"} 4
xinference:requests_total_counter{method="None",path="/qwen-chat"} 1
xinference:requests_total_counter{method="GET",path="/v1/cluster/auth"} 1
xinference:requests_total_counter{method="GET",path="/v1/models/{model_uid}"} 1
xinference:requests_total_counter{method="POST",path="/v1/chat/completions"} 1
# HELP xinference:responses_total_counter Total number of responses sent.
# TYPE xinference:responses_total_counter counter
xinference:responses_total_counter{method="GET",path="/v1/model_registrations/{model_type}"} 1
xinference:responses_total_counter{method="GET",path="/v1/cluster/devices"} 1
xinference:responses_total_counter{method="GET",path="/ui"} 1
xinference:responses_total_counter{method="POST",path="/v1/models"} 1
xinference:responses_total_counter{method="GET",path="/v1/models/"} 2
xinference:responses_total_counter{method="GET",path="/v1/models"} 2
xinference:responses_total_counter{method="HEAD",path="/qwen-chat"} 1
xinference:responses_total_counter{method="POST",path="/v1/ui/{model_uid}"} 1
xinference:responses_total_counter{method="GET",path="/qwen-chat"} 4
xinference:responses_total_counter{method="POST",path="/qwen-chat"} 4
xinference:responses_total_counter{method="GET",path="/v1/cluster/auth"} 1
xinference:responses_total_counter{method="GET",path="/v1/models/{model_uid}"} 1
xinference:responses_total_counter{method="POST",path="/v1/chat/completions"} 1
# HELP xinference:status_codes_counter Total number of response status codes.
# TYPE xinference:status_codes_counter counter
xinference:status_codes_counter{method="GET",path="/v1/model_registrations/{model_type}",status_code="200"} 1
xinference:status_codes_counter{method="GET",path="/v1/cluster/devices",status_code="200"} 1
xinference:status_codes_counter{method="GET",path="/ui",status_code="404"} 1
xinference:status_codes_counter{method="POST",path="/v1/models",status_code="200"} 1
xinference:status_codes_counter{method="GET",path="/v1/models/",status_code="307"} 2
xinference:status_codes_counter{method="GET",path="/v1/models",status_code="200"} 2
xinference:status_codes_counter{method="HEAD",path="/qwen-chat",status_code="404"} 1
xinference:status_codes_counter{method="POST",path="/v1/ui/{model_uid}",status_code="200"} 1
xinference:status_codes_counter{method="GET",path="/qwen-chat",status_code="307"} 1
xinference:status_codes_counter{method="GET",path="/qwen-chat",status_code="200"} 3
xinference:status_codes_counter{method="POST",path="/qwen-chat",status_code="200"} 4
xinference:status_codes_counter{method="GET",path="/v1/cluster/auth",status_code="200"} 1
xinference:status_codes_counter{method="GET",path="/v1/models/{model_uid}",status_code="200"} 1
xinference:status_codes_counter{method="POST",path="/v1/chat/completions",status_code="200"} 1
# HELP xinference:time_to_first_token_ms First token latency in ms.
# TYPE xinference:time_to_first_token_ms gauge
xinference:time_to_first_token_ms{format="pytorch",model="qwen-chat",node="127.0.0.1:47981",quantization="none",type="LLM"} 20076.820135116577