Class: Tep::Llm::OpenAI::Backend

Inherits:
Object
  • Object
show all
Defined in:
lib/tep/openai_server.rb

Overview

The interface an app's backend implements. Defaults make a bare backend safe to compile + serve (empty model list, chat unsupported, cpu device). Subclasses override what they offer.

Instance Method Summary collapse

Instance Method Details

#chat_completion(req) ⇒ Object

Message-level (chat) generation. Mirrors generate_from_tokens but receives the raw req so the backend can parse the messages array itself + apply its own chat template. Tep doesn't pre-build a Message because templating + role ordering is per-model; the JSON tools live in SpinelKit::Json. The return is reused from the token path (text becomes the assistant message's content). Base no-op; subclasses override. Only reached when supports_chat? returns true -- the handler gates with a 501 otherwise.



77
78
79
 
# File 'lib/tep/openai_server.rb', line 77

def chat_completion(req)
  Tep::Llm::OpenAI::Completion.new
end

#chat_completion_stream(req, sink) ⇒ Object

Streaming chat (#127). Per-token variant for SSE /v1/chat/completions when the request carries "stream":true. Backend writes each token to sink via sink.emit_token(piece); the sink formats it as the OpenAI chat-streaming delta frame and writes one chunked frame. Same subclass-override-sink pattern as 7.2 (generate_stream_from_tokens). Base no-op.



87
88
89
 
# File 'lib/tep/openai_server.rb', line 87

def chat_completion_stream(req, sink)
  0
end

#device_kind ⇒ Object

Backend's device, surfaced into the run_start event's backend.kind at serve! time. Defaults to cpu.



93
94
95
 
# File 'lib/tep/openai_server.rb', line 93

def device_kind
  "cpu"
end

#generate_embeddings(model, token_ids) ⇒ Object

Embedding generation for /v1/embeddings. token_ids is the encoded input (Array; this server speaks IDs only, tokenize client-side, same policy as generate_from_tokens). Returns the pooled embedding as an Array of length d_model -- the backend owns the lookup + pooling strategy (toy mean-pools per-token embeddings). Base returns an empty vector so a bare backend compiles; only reached when supports_embeddings? is true (EmbeddingsHandler gates 501).



118
119
120
121
122
 
# File 'lib/tep/openai_server.rb', line 118

def generate_embeddings(model, token_ids)
  empty = [0.0]
  empty.delete_at(0)
  empty
end

#generate_from_tokens(model, token_ids, sampling) ⇒ Object

PRIMARY shape: token-level generation (maps to /v1/completions, non-streaming). token_ids is the encoded prompt (Array); sampling is a Tep::Llm::OpenAI::Sampling. Returns a Tep::Llm::OpenAI::Completion (text + usage). The base returns an empty completion so a bare backend compiles; real backends override.



44
45
46
 
# File 'lib/tep/openai_server.rb', line 44

def generate_from_tokens(model, token_ids, sampling)
  Tep::Llm::OpenAI::Completion.new
end

#generate_stream_from_tokens(model, token_ids, sampling, sink) ⇒ Object

STREAMING shape (7.2): the per-token variant for SSE /v1/completions when the request carries "stream": true. The backend writes each token to sink via sink.emit_token(piece); the sink (Tep::Llm::OpenAI::StreamSink) formats it as an OpenAI SSE frame and writes to the outbound chunked stream. Blocks/yields don't lower across the spinel boundary, so a typed sink replaces the block -- backends never see SSE wire format or the client fd. Base no-op (subclasses override).



57
58
59
 
# File 'lib/tep/openai_server.rb', line 57

def generate_stream_from_tokens(model, token_ids, sampling, sink)
  0
end

#list_models ⇒ Object

Available model names -> [String]. /v1/models wraps these.



31
32
33
34
35
 
# File 'lib/tep/openai_server.rb', line 31

def list_models
  empty = [""]
  empty.delete_at(0)
  empty
end

#model_owner ⇒ Object

owned_by value for each entry in the /v1/models list. Defaults to "tep"; a backend overrides to attribute models to its own project (e.g. toy returns "toy").



100
101
102
 
# File 'lib/tep/openai_server.rb', line 100

def model_owner
  "tep"
end

#supports_chat? ⇒ Boolean

Does this backend implement message-level (chat) generation? When false, /v1/chat/completions returns 501. (The chat template is per-model + an ML concern; tep doesn't ship one.)

Returns:

  • (Boolean) 


64
65
66
 
# File 'lib/tep/openai_server.rb', line 64

def supports_chat?
  false
end

#supports_embeddings? ⇒ Boolean

Backends that can embed override this -> true (gates /v1/embeddings, chunk 7.3).

Returns:

  • (Boolean) 


106
107
108
 
# File 'lib/tep/openai_server.rb', line 106

def supports_embeddings?
  false
end