Make HOURS_VALUES a host array to avoid import-time GPU preallocation#13
Open
hmgaudecker wants to merge 9 commits into
Open
Make HOURS_VALUES a host array to avoid import-time GPU preallocation#13hmgaudecker wants to merge 9 commits into
hmgaudecker wants to merge 9 commits into
Conversation
The module-level `HOURS_VALUES = jnp.array(...)` materialized on the default device at import. With XLA_PYTHON_CLIENT_PREALLOCATE=true that first array op reserves 95% of device 0 in every process that imports the model — including the MSM estimation's pytask orchestrator, which only `srun`s GPU ranks and must leave the devices free for them. The orchestrator thus starved the rank's pool reservation, surfacing as the device-0 OOM. Make HOURS_VALUES a host (NumPy) array, converted to JAX at the indexing sites where the value folds into the surrounding compiled function. No numerical change. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ery state The three leisure functions returned a raw `time_endowment - losses` with no floor, so once work costs reached the endowment leisure went to zero or negative and fed a non-positive base into the CRRA aggregator (NaN utility), and the kink made the MSM objective non-smooth for the derivative-free optimizer. A shared `_smooth_leisure_floor` helper applies a scaled softplus (`smoothing * logaddexp(0, available / smoothing)`, smoothing = 1% of the endowment) so leisure bends smoothly to 0+ instead. It reduces to `available` in the bulk, so existing estimates are preserved; the fixed-cost/reentry parameters stay identified when the optimizer drives them high. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The softplus leisure floor keeps leisure strictly positive in every state-action cell, so the `positive_leisure` feasibility constraint never binds. Remove it from the canwork retiree/nongroup/tied regime builders and delete the unused `positive_leisure` helper; feasibility is now carried by the smooth floor in the leisure functions themselves. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
BQSEGM solves one 1-D consumption/savings regime with at most one discrete action, so unlike DC-EGM it attaches per regime: solver="bqsegm" gives the M1 slice regime nongroup_nomc_inelig_canwork a BQSEGM config (budget node `resources`, post-decision `savings`, the savings form shared with DC-EGM) and leaves every other living regime on brute force. The three savings-form gates in _common accept "bqsegm" alongside "dcegm"; the HIS builders attach whichever EGM solver is present. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The BQSEGM case-piece envelope handles at most one discrete action, so the M1 vertical slice drops buy_private as a choice and binds it to BuyPrivate.yes in its consumers (premium, OOP) via the dags remove-and-fix convention. The brute and DC-EGM paths keep buy_private as an action. The drop_labor_supply hook is added for the next step (the pure-continuous slice also fixes labor_supply). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The pure-continuous M1 slice drops labor_supply as a discrete action and supplies it as a fixed full-time node (LaborSupply.h2000) read by labor income, AIME accrual, and the lagged-supply transition. lagged_labor_supply stays a state so the cross-regime continuation space is unchanged; the nomc regime's lifecycle transition is unaffected (it never gated on labor). With buy_private already fixed, the M1 regime now has no discrete action, so the only choice is continuous consumption against the cliffed budget. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…tax cliff The M1 vertical-slice regime `nongroup_nomc_inelig_canwork` carries `assets` (the Euler axis) alongside `aime` and the stochastic shock grids, so the BQSEGM config names `continuous_state="assets"`; the rest ride along. The progressive federal income tax is declared as a piecewise-affine schedule on `gross_income`, kinking at each finite bracket edge `income_tax_schedule.brackets_upper[ spousal_income, k]`, so BQSEGM differentiates the budget per declared bracket. The decorator is metadata-only — brute and DC-EGM solve identically. With these the full model constructs under `solver="bqsegm"` (M1 via BQSEGM, every other living regime on brute force). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
DC-EGM solves every living regime, so its solver-contract functions and the dropped borrowing constraint are broadcast model-wide. BQSEGM solves only the M1 regime, so broadcasting them forced every brute regime to supply the solver's `marginal_continuation` and stripped their borrowing constraint. Keep the model-level broadcast for DC-EGM only. The M1 regime carries the savings-form budget functions (`resources`, `savings`) at regime level and masks the borrowing constraint — BQSEGM enforces the borrowing limit through its savings grid's lower bound. BQSEGM inverts the Euler equation internally (CRRA from the utility parameters), so unlike DC-EGM it does not carry `inverse_marginal_utility`; a new `build_bqsegm_functions` supplies just the budget pair. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The BQSEGM continuation integrates the child's stochastic next-states over their joint node mesh; reading it in one pass scales the peak intermediate with the full ride-along x node x child-grid product, which on the M1 slice (aime=38 fixed PIA-bend grid x many ride-along axes) is enormous. Expose n_bqsegm_stochastic_node_batch_size on GridConfig and pass it into the solver, so a memory-rich device reads the whole mesh in one pass (0, default) while a tighter budget loops it in blocks. Resolves the M1 solve OOM together with the upstream pylcm splay threading. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
HOURS_VALUES = jnp.array(...)was a module-level JAX array: importing the model materialized it on the default device, and underXLA_PYTHON_CLIENT_PREALLOCATE=truethat first array op reserved 95% of device 0 in every importing process — including the MSM estimation's pytask orchestrator, which onlysruns GPU ranks and must leave the devices free. The orchestrator thus starved the rank's pool reservation (the device-0 OOM). Made it a host (NumPy) array, converted to JAX at the indexing sites. No numerical change.🤖 Generated with Claude Code