Skip to content

python_torch_trace_provider.md

Latest commit

 

History

History
81 lines (63 loc) · 2.18 KB

python_torch_trace_provider.md

File metadata and controls

81 lines (63 loc) · 2.18 KB

Python Torch Trace Provider

Purpose: emit Crucible trace events from real Hugging Face Transformers models through PyTorch when Elixir-native providers cannot expose required internals.

What It Captures

tools/python/crucible_torch_trace.py runs a real AutoModelForCausalLM forward pass and a manual greedy generation loop. It writes bounded summaries, not raw tensor arrays.

Captured signals:

  • input_ids
  • attention_mask
  • final_logits
  • hidden_state
  • attention_weights
  • generation_step_logits
  • generation_token
  • decoded_text

The worker writes:

  • tmp/crucible_v5/traces/python/<name>.trace.jsonl
  • tmp/crucible_v5/capability_reports/<name>.capability_report.json
  • tmp/crucible_v5/reports/<name>.python_trace_report.json

Commands

Run the tiny GPT-2 control:

python3 tools/python/crucible_torch_trace.py \
  --model-id hf-internal-testing/tiny-random-gpt2 \
  --model-family gpt2 \
  --prompt "Hi" \
  --artifact-root tmp/crucible_v5 \
  --trace-name python_torch_tiny_random_gpt2_phase15 \
  --max-new-tokens 4 \
  --top-k 5

Run standard GPT-2:

python3 tools/python/crucible_torch_trace.py \
  --model-id gpt2 \
  --model-family gpt2 \
  --prompt "Hi" \
  --artifact-root tmp/crucible_v5 \
  --trace-name python_torch_gpt2_phase15 \
  --max-new-tokens 3 \
  --top-k 5

Replay through Trinity:

mix trinity.crucible.inspect \
  --trace tmp/crucible_v5/traces/python/python_torch_gpt2_phase15.trace.jsonl \
  --artifact-root tmp/crucible_v5

mix trinity.crucible.matrix_eval \
  --trace tmp/crucible_v5/traces/python \
  --artifact-root tmp/crucible_v5

Boundary Decision

For V5, this worker is an external trace provider, not a supervised Python runtime. snakepit is the candidate long-lived Python process pool and snakebridge is the candidate higher-level Python binding/contract layer. They are not inserted into this phase because the acceptance gate is trace proof over real model internals, and the stable contract is the Crucible JSONL trace plus capability report.

The next supervised provider phase should put Python worker lifecycle under self_hosted_inference_core provider behavior after the trace contract has stayed stable.