MCPcopy Create free account
hub / github.com/RL-Align/RL-Kernel / configure_stateless_model

Function configure_stateless_model

rl_engine/executors/stateless_executor.py:335–366  ·  view source on GitHub ↗

Apply best-effort no-cache scoring knobs without importing optional runtimes. Hugging Face-style models decide attention kernels from their config rather than forward kwargs. Setting these attributes keeps the wrapper explicit while still allowing plain PyTorch modules to run uncha

(
    model: torch.nn.Module,
    config: StatelessForwardConfig,
)

Source from the content-addressed store, hash-verified

333 "active_completion_tokens": active_tokens,
334 "device": str(input_ids.device),
335 "dtype": str(input_ids.dtype).replace("torch.", ""),
336 "elapsed_ms": elapsed_seconds * 1000.0,
337 "use_cache": False,
338 "use_cache_passed": bool(use_cache_passed),
339 "detached_outputs": bool(detached_outputs),
340 "zero_kv_cache": kv_cache_summary.tensor_count == 0,
341 "kv_cache_output_present": kv_cache_summary.tensor_count > 0,
342 "kv_cache_output_tensors": kv_cache_summary.tensor_count,
343 "kv_cache_output_bytes": kv_cache_summary.total_bytes,
344 "kv_cache_output_mb": kv_cache_summary.total_mb,
345 **dict(no_cache_policy),
346 }
347 if cuda_tracking:
348 device = input_ids.device
349 metrics["peak_allocated_mb"] = torch.cuda.max_memory_allocated(device) / 1_048_576.0
350 metrics["peak_reserved_mb"] = torch.cuda.max_memory_reserved(device) / 1_048_576.0
351 return metrics
352
353
354def configure_stateless_model(
355 model: torch.nn.Module,
356 config: StatelessForwardConfig,
357) -> dict[str, float | int | str | bool]:
358 """
359 Apply best-effort no-cache scoring knobs without importing optional runtimes.
360
361 Hugging Face-style models decide attention kernels from their config rather
362 than forward kwargs. Setting these attributes keeps the wrapper explicit
363 while still allowing plain PyTorch modules to run unchanged.
364 """
365
366 use_cache_targets = 0
367 attention_targets = 0
368
369 for target in _model_config_targets(model):

Calls 1

_model_config_targetsFunction · 0.85

Tested by

no test coverage detected