dasLLAMA: the engine refuses a bare job queue, and every queue in the tree is configured

[borisbat]

Behavior change: a dasLLAMA program that opens with_job_que() without setup_dasllama_jobque() now panics at its first engine dispatch. DASLLAMA_ALLOW_BARE_JOBQUE=1 keeps it running with a warning.

Why. A queue nobody configured clones a fork context per job and parks each worker after it. The engine issues ~160 dispatches per decoded token, so an E2B q8 forward step took 3.6 s on such a queue against 16 ms configured. test_ple_modes alone ran 485 s on this, and half the stocked suite's wall was the same regime.

What changes.

• get_jobque_fork_pool() (jobque module) reads the per-context pool flag. It is the one knob setup_dasllama_jobque() always sets, nothing else sets it, and it outlives the queue.
• The first counted dispatch of a process checks it once (check_jobque_configured in dasllama_par.das). A queue without the pool panics naming the setup call; the env knob turns the panic into a warning. No queue at all is not checked.
• The loader's own scoped queues (load_gguf, load_gguf_streaming) configure themselves. Every bare block in the tree gets the setup call: dasllama-convert (2), six test files, two matmul benches, eight harness probes.
• skills/internal/make_pr.md: the PR body's top layer is four labeled parts (Why, What changes, Observable behavior, Where to look), ~100 words, bullets only.

Observable behavior.

• Bare-queue generate: 200x slower, silent -> panic at the first dispatch, message names the fix.
• test_ple_modes: 485 s -> 34 s. test_batch_decode: 264 s -> under 30 s. Stocked suite: 1577 s -> 1018 s.
• dasllama-convert: conversion passes ran on a bare queue -> configured queue.

Where to look. check_jobque_configured in dasllama_par.das; the two loader blocks in dasllama_load.das; ARCHITECTURE_RUNTIME.md section 2.20.

Validation, claims, ledger

Validation

• Stocked suite on the M5 box: 39 files, 0 failed, 1018 s (baseline 1577 s). Model-free: 65 files, 0 failed.
• The large-tier arms (DASLLAMA_PARITY_FULL=1) did not run; their wall under the tripwire is unmeasured.
• preflight --full green after one fix: the new builtin joined the jobque page's "Internal invocations" group in das2rst.

Claims - stated, not tested

• The per-knob attribution of the bare regime (team mode alone leaves 2.0 s, each other knob alone ~25 ms) is a probe measurement on E2B q8. A break would show as a stocked-suite wall regression, not a test failure.
• The harness probes and bench_matmul_par that gained the setup call compile on all lint rails but were not run.

Not done

• A DASLLAMA_PARITY_FULL=1 stocked run under the tripwire.
• The ~13 s per-file floor of the model suites: dastest's runtime compile_file runs past the AST module cache.

🤖 Generated with Claude Code

[Copilot]

🟡 Changes recommended

The new tripwire test’s subprocess command construction and the user-facing panic/doc wording need small adjustments to avoid flakiness and misleading guidance.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR makes dasLLAMA refuse “bare” job queues (created via with_job_que() but not configured via setup_dasllama_jobque()), because that regime is catastrophically slow for the engine’s fork/join dispatch; it adds an environment-variable escape hatch, and updates in-tree call sites so existing programs/tests/benches run on configured queues by default.

Changes:

• Add get_jobque_fork_pool() to detect whether the current context is configured for fork-context pooling, and use it to tripwire bare-queue engine dispatch.
• Introduce DASLLAMA_ALLOW_BARE_JOBQUE to downgrade the tripwire from panic to warning, and document the knob.
• Update loaders/tools/tests/benches/harness probes to call setup_dasllama_jobque*() inside with_job_que() blocks; add a model-free tripwire test.

File summaries

File	Description
utils/dasllama-convert/main.das	Configures the job queue in tool-local with_job_que() blocks to avoid bare-queue dispatch.
src/builtin/module_builtin_jobque.cpp	Adds get_jobque_fork_pool builtin and binds it into the jobque module.
include/daScript/simulate/aot_builtin_jobque.h	Exposes the new get_jobque_fork_pool symbol to AOT-generated code.
doc/source/stdlib/handmade/function-jobque-get_jobque_fork_pool-0x24aa9e9140117fba.rst	Documents the new jobque query builtin.
doc/reflections/das2rst.das	Classifies get_jobque_fork_pool in the das2rst jobque docs grouping.
modules/dasLLAMA/dasllama/dasllama_par.das	Implements the bare-queue tripwire (panic/warn) at first counted dispatch.
modules/dasLLAMA/dasllama/dasllama_env.das	Adds allow_bare_jobque to the EngineEnv environment configuration.
modules/dasLLAMA/ENVIRONMENT.md	Documents DASLLAMA_ALLOW_BARE_JOBQUE.
modules/dasLLAMA/dasllama/dasllama_load.das	Ensures internally-spawned loader job queues are configured.
modules/dasLLAMA/dasllama/dasllama.das	Updates public API docs to mention the new job-queue configuration requirement.
modules/dasLLAMA/tests/test_ple_modes.das	Configures job queue before engine use in this test’s with_job_que() block.
modules/dasLLAMA/tests/test_forward.das	Configures job queue before engine use across multiple with_job_que() blocks.
modules/dasLLAMA/tests/test_batch_decode.das	Configures job queue before engine use across multiple with_job_que() blocks.
modules/dasLLAMA/tests/test_parity_pregate.das	Ensures the slice cell’s with_job_que() configures the queue; adds required module.
modules/dasLLAMA/tests/test_model_image_vulkan.das	Configures job queue for the cached-load with_job_que() block.
modules/dasLLAMA/tests/test_jobque_tripwire.das	Adds a model-free test that asserts panic/warn/silent behavior for bare vs configured queues.
modules/dasLLAMA/tests/_jobque_tripwire_root.das	Adds the spawned fixture program used by the tripwire test.
modules/dasLLAMA/tests/run.das	Registers the new tripwire test in suite selection lists.
modules/dasLLAMA/tests/CLAUDE.md	Documents the new tripwire test and the “always setup inside with_job_que” rule for tests.
modules/dasLLAMA/benchmarks/matmul/test_matmul_par.das	Configures job queue for threaded matmul tests.
modules/dasLLAMA/benchmarks/matmul/bench_matmul_par.das	Configures job queue for the matmul benchmark driver.
modules/dasLLAMA/harness/tq4_order_probe.das	Configures job queue in the probe’s with_job_que() block.
modules/dasLLAMA/harness/tq4_basis_probe.das	Configures job queue in the probe’s with_job_que() blocks.
modules/dasLLAMA/harness/residency_ramp_probe.das	Requires the setup helper and configures the job queue for the probe.
modules/dasLLAMA/harness/quant_eval.das	Requires the setup helper and configures the job queue for the harness run.
modules/dasLLAMA/harness/quant_eval_q4.das	Requires the setup helper and configures the job queue for the harness run.
modules/dasLLAMA/harness/gen_tune_probe.das	Configures job queue before using job-queue-dependent tuning/dispatch behavior.
modules/dasLLAMA/ARCHITECTURE_RUNTIME.md	Documents the new “configured queue only” policy and its performance rationale.
skills/internal/make_pr.md	Updates PR-body format guidance (Why/What changes/Observable behavior/Where to look).

Review details

• Files reviewed: 29/29 changed files
• Comments generated: 4
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🟡 Changes recommended

The new diagnostics/docs contain a couple of concrete accuracy issues (wording/claims) that should be corrected to match real call paths and existing configuration sites.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (2)

modules/dasLLAMA/dasllama/dasllama_par.das:63

• The panic/warning text says to call setup_dasllama_jobque() before the first generate/eval, but this tripwire can also fire during other engine dispatches (e.g., load/conversion work that uses maybe_parallel_for). Widening the wording would make the diagnostic accurate regardless of which call triggers the first dispatch.

    let msg = "dasLLAMA: the job queue is not configured for the engine's fork/join dispatch - call setup_dasllama_jobque() inside with_job_que() before the first generate/eval (a bare queue decodes ~200x slower; DASLLAMA_ALLOW_BARE_JOBQUE=1 runs anyway)"

modules/dasLLAMA/dasllama/dasllama.das:15

• This docstring says the setup call is needed before the first generate/eval, but the job-queue tripwire can also fire during load_model (dasllama_load.das uses maybe_parallel_for during conversion), so callers wrapping load in with_job_que() will hit the panic earlier than this text implies.

//! call ``setup_dasllama_jobque()`` inside it before the first ``generate``/``eval`` — the engine

• Files reviewed: 29/29 changed files
• Comments generated: 1
• Review effort level: Lite

[Copilot]

🟢 Approval recommended

The behavior change is guarded, well-documented, and backed by a focused regression test, with only minor doc-comment readability feedback.

Review details

• Files reviewed: 29/29 changed files
• Comments generated: 1
• Review effort level: Lite

[Copilot]

🟢 Approval recommended

The behavior change is covered by a dedicated tripwire test and systematic call-site updates, with only minor documentation-consistency nits remaining.

Review details

• Files reviewed: 29/29 changed files
• Comments generated: 2
• Review effort level: Lite