[joelteply]

ggml-opt: freeing an optimizer checks only its own graph, so an idle shadow unloads while another mind learns

c5551ed made ggml_opt_free assert that the SHARED scheduler had no pending graph. A serving
context keeps one optimizer per shadow on one scheduler, and llama_context::opt_release frees an
idle shadow's optimizer while another shadow's step is in flight (#85's contract: "ggml_opt_free
touches no shared scheduler"). That assert then aborted llama-server whenever the unload landed
mid-step (fork #106/#107 reproducer, learn-unload-during-learn.py, 2 of 3 runs on CPU).

ggml_opt_free releases only this context's own buffers and metadata. The check is now
context-local: ggml_backend_sched_graph_pending_for(sched, graph), true only when the scheduler is
part-way through THAT graph (graph_step already keys its continuation on the caller's graph
pointer). Freeing a context whose own graph is in flight still aborts. ggml_opt_reset and
ggml_opt_prepare_alloc keep the scheduler-wide guard: reset zeroes the last graph's gradients,
which in a dynamic graph can sit in the shared compute buffer another mind now uses, and
prepare_alloc allocates onto the scheduler.

Tests (test-opt-dynamic-accum, built with -fsanitize=address on CPU, IntelMac):

• run_frees_idle_mind_while_another_steps: B learns a step, A stops one node into its step
  (scheduler pending), B is freed, A completes with exactly the update it makes alone. PASS.
• run_free_own_pending_graph_aborts: freeing a context whose own graph is in flight aborts
  (observed in a forked child; skipped with a message where fork() is unavailable). PASS.
• the existing suite: PASS.
• Mutation: restoring the scheduler-wide check aborts the first test at the same assert the
  engine hit (exit 134).
  End to end: the reproducer passes 10/10 on the fixed CPU server, but it is timing-dependent (it
  aborts only if the unload lands mid-graph; 0/6 on the unfixed engine in a later run), so the unit
  test is the proof.

Card 5cfd134a. Base: codex/s4-per-mind-integration.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FRQWzgSfo79JtnwZywHKrE

[joelteply]

Review (Fable): PASS. The claim checked: the narrowed assert still guards a context's OWN in-flight step. ggml_opt_eval_step steps the scheduler with exactly opt_ctx->allocated_graph_copy (ggml-opt.cpp:2039), and the scheduler records that pointer as continuation_graph (ggml-backend.cpp:2742), so ggml_backend_sched_graph_pending_for(sched, allocated_graph_copy) is true exactly when this context's graph is mid-step; another context's pending graph no longer blocks freeing an idle one. The forked-child test pins the abort for the own-pending case.