mtplx 2.12.1__tar.gz → 2.12.2__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {mtplx-2.12.1 → mtplx-2.12.2}/CHANGELOG.md +14 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/PKG-INFO +1 -1
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/openai.py +149 -5
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/version.py +2 -2
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx.egg-info/PKG-INFO +1 -1
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx.egg-info/SOURCES.txt +1 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/pyproject.toml +1 -1
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_host_memory_allowance.py +62 -7
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memguard_admission.py +30 -26
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memguard_backend_contract.py +14 -4
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memguard_limit_allowances.py +3 -1
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prompt_scoring_admission.py +8 -3
- mtplx-2.12.2/tests/test_small_mac_prefill.py +283 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/CITATION.cff +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/LICENSE +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/MANIFEST.in +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/NOTICE +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/README.md +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/a3b_compiled_target_prefix.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/a3b_mtp_batch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/a3b_whole_moe.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/adaptive.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/adaptive_dtemp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/app_settings.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/aria2_downloader.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/arrays_cache_patch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/artifacts.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/attention_context.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/attention_math.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/attention_split.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/backends/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/backends/deepseek_mtp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/backends/descriptors.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/backends/family_settings.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/backends/gemma4_assistant.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/backends/glm_mtp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/backends/hy_v3_mtp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/backends/mimo_mtp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/backends/nemotron_h_mtp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/backends/qwen3_next.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/backends/registry.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/backends/step3p5_mtp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/batched_decode.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/batching/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/batching/admission.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/batching/buckets.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/batching/scheduler.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/batching/state.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/code_eval.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/programming_prompts.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/prompts/aime_2026.jsonl +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/prompts/calibration_coding.jsonl +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/prompts/default.jsonl +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/prompts/flappy.jsonl +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/prompts/long_code.jsonl +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/prompts/long_code_uncapped.jsonl +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/prompts/python_modules_long.jsonl +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/aime.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/batch_equivalence.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/capture_commit_equivalence.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/competitor_baselines.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/contract_probe.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/harness.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/mtp1_gate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/mtp1_sampler_smoke.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/mtp_adaptive.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/mtp_chain_probe.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/mtp_depth_grid.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/mtp_depth_sweep.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/mtp_tree_probe.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/multi_qmv_probe.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/preflight.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/runtime_smoke.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/session_bank.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/truth.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/verify_profile.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/verify_qmm_probe.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/runners/verify_ratio.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/schema.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/validators/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/validators/aime.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/benchmarks/validators/basic.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/block_attention.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/cache_bank/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/cache_bank/codec.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/cache_bank/cold_tier.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/cache_bank/disk_budget.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/cache_bank/reconcile.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/cache_state.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/chat_encode_cache.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/chat_encoding.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/checkpoint_anchors.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/cli.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/commands/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/commands/forge.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/commands/forge_mixed_convert.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/commands/forge_qwen4_audit.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/commands/forge_qwen4_exp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/commands/public.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/commands/trace.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/commands/trace_analysis.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/commands/trace_clients.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/commands/trace_metrics.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/commands/trace_record.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/commands/trace_report.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/compile_state.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/compiled_forward.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/compressed_tensors.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/config.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/constants.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/constrained.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/context_copy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/correctors/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/correctors/diagonal_affine.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/correctors/low_rank.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/daemon_client.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/dashboard/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/dashboard/_static/assets/index-BTDFdIWu.css +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/dashboard/_static/assets/index-DIdoATFR.js +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/dashboard/_static/index.html +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/data/qwen38_code_ranked_64k.npy +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/deepseek_mtp_patch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/deepseek_v4_adaptive_width.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/deepseek_v4_attention_island.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/deepseek_v4_attn_proj_wide_m3.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/default_models.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/demotions.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/dense_mrope.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/diagnostics.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/draft_lm_head.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/draft_sampling.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/engine_session.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/env.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/errors.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/expert_layout.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/fan_mode.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/fast_sampling.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/float32_operand.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/forkev_telemetry.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/frspec_draft.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/gdn_capture.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/gemma4_pair.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/generation.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/glm_mtp_patch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/graphbank.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/hardware.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/hf_loader.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/hy_v3_mtp_patch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/jsonc.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernel_selfcheck.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/a3b_whole_moe.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/copy_leaf.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/fused_norm.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/gdn_blocked_prefill.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/gdn_conv_norm.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/gdn_gated_norm.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/gdn_out_fused.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/gdn_prefill_prework.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/gdn_step_fused.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/gdn_tape_headquarter.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/hadamard_rotate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/hc_prefill.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/hc_verify_read.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/hyper_connection.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/hyper_connection_v3.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_decode.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_dense_mlp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_gated_oproj.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_moe_combine.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_moe_gather_gemm.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_moe_merged.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_moe_shared.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_moe_swiglu.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_prefill_moe_combine.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_prefill_qk_rope.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_prefill_router.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_qk_rope_sliding.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_qk_yarn_full.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_qkvg_fused.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_residual_router.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_route_csort.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_router_topk.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_sdpa_2pass.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_sdpa_pair.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/laguna_steel_attn.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/lm_head_topk.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/logits_topk.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/moe_glu_decode.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/moe_sorted_gather_nax.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/native_gdn_tail.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qmv4_matmul.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qsa_dense_band_sdpa.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qsa_flash_skip.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qsa_indexer_compile.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qsa_indexer_prefill.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qsa_indexer_prepare.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qsa_indexer_select.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qsa_pooled_row.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qsa_prefill_direct.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qsa_prefill_flash.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qsa_prefill_probe.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qsa_verify_select.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qwen4_m4_rope.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qwen4_m4_route.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qwen4_m4_routed_down.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qwen4_m4_routed_glu.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qwen4_m4_stage3.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qwen4_moe_prefill_combine.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/qwen4_qsa_m4_fused_kv_gather.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/sdpa_2pass.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/sdpa_2pass_paged.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/sdpa_2pass_paged_q8.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/sdpa_gqa_packed.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/sdpa_gqa_packed_quant.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/sdpa_nax_flash.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/sdpa_nax_flash_dsplit.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/sdpa_nax_tile.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/ternary_qmv.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/verify_mlp_fused.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kernels/vqmm_crossrow.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kpi/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kpi/reference_vllm.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kpi/runtime_kpis.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/kv_quant.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/laguna_alt_step.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/laguna_compiled_step.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/lane_explain.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/launch_lane.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/lfm2_fast.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/loop_guard.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/memory_plan.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/metadata_scrub.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/mimo_mtp_patch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/mlx_process_env.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/model_catalog.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/model_scheduler.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/model_updates.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/models/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/models/deepseek_v4.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/models/laguna.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/models/laguna_config.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/models/laguna_fused.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/models/prism_hadamard_qwen35.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/models/qwen4_exp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/moe_packed_projections.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/moe_sorted_gather.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/mtp_activation_stats.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/mtp_adapters.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/mtp_batch_numerics.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/mtp_history_cache_only.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/mtp_patch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/native_mlp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/nax_detect.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/nax_verify.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/nemotron_h_mtp_patch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/ngram_row_layout.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/one_copy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/opencode.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/optimization_profiles.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/os_memory.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/packed_concats.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/pi.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/ple_prefill_lookahead.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/ple_row_gather.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/prefill_bench.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/prefill_plan.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/prefill_rungs.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/profiles.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/progress_heartbeat.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/proj_fusion.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/proj_quant.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/proposal_reranker.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/qsa_mtp_precompute.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/qsa_restore_staging.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/qwen3_5_mtp_patch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/qwen4_block_verify.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/qwen4_claim_contract.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/qwen4_draft_device_chain.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/qwen4_draft_k20_prescatter.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/qwen4_fixed_verify.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/qwen4_m4_stage3.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/qwen4_prefill_profile.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/qwen4_verify_glue.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/qwen_row_owned_router.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/ragged_attention.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/ragged_kv_cache.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/reasoning_codecs.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/reasoning_effort.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/request_capture.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/retrieval.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/roofline_profile.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/rope_origin.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/route_tape.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/runtime.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/runtime_options.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/sampling.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/dashboard_state.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/flight_recorder.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/hyper.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/mtp_batch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/omlx_bridge/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/omlx_bridge/adapter.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/omlx_bridge/thinking.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/omlx_bridge/tool_calling.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/prefill_safety.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/request_policy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/request_spans.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/response_envelope.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/responses.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/retry_prompt.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/served_window.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/stream_rate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server/stream_recovery.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/server_urls.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/session_bank.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/speculative.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/step3p5_mtp_patch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/swival.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/system_memory.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/templates/qwen36_froggeric_v19/chat_template.jinja +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/templates/qwen36_froggeric_v21_3/chat_template.jinja +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/thermal.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/thermal_pressure.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/thermal_sidecar.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/thinking_guard.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/trace_parity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/turboquant.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/ui/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/ui/banner.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/ui/chat_printer.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/ui/download_progress.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/ui/onboarding.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/ui/panels.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/ui/progress.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/vendored_arrays_cache.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/vendored_hy_v3.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/verify_crossrow.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/verify_kernels.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/verify_qmv.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/vision/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/vision/mrope.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/vision/processing.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/vision/qwen3_vl_tower.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/vision/splice.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx/vision_graft.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx.egg-info/dependency_links.txt +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx.egg-info/entry_points.txt +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx.egg-info/requires.txt +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/mtplx.egg-info/top_level.txt +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/agent_session_gate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/agent_user_path_qa.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/aime_serve_gate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/aime_shape_memory_bench.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/analyze_head_sweeps.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/audit_agent_prompt_fidelity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/audit_catalog_sizes.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/bonsai_memory_table.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/build_bonsai_mtplx_pack.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/build_flash_next_quality_pack.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/build_flash_next_quality_pack.sh +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/bundle_native_runtime_wheel.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/check_ai_attribution.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/check_prompt_scoring.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/code_eval_gate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/collect_mtp_activation_stats.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/collect_mtp_hidden_calib.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/combine_mtp_adapters.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/compiled_verify_exactness.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/convert_qwen4_exp_fp8.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/convert_step37_to_mtplx_step3p5.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/decodecliff_probe_20260830.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/decodecliff_pulse_cell_20260830.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_adaptive_width_arms.sh +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_adaptive_width_guarded.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_attention_island_arms.sh +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_attention_island_guarded.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_attn_proj_wide_m3_arms.sh +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_attn_proj_wide_m3_guarded.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_build_mtp_model.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_decode_verify.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_dispatch_census.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_guard_window.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_logits_gate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_moe_tail_arms.sh +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_moe_tail_gate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_moe_tail_guarded_bracket.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_mtp_bind_check.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_mtpk_bench.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_op_census.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_smoke_generate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_sparse_gate_probe.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/deepseek_v4_validate_moe_tail_k3_bracket.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/dense_mrope_parity_ours.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/dense_mrope_reference_check.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/dense_vision_compiled_parity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/eval_mtp_corrector.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/filter_mtp_adapter.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/fp16_turbo_exactness_20260707.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/fresh_venv_smoke.sh +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/gauntlet_scoreboard.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/gen_models_manifest.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/graft_vision_tower.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/hygiene_scan.sh +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/install_macos.sh +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/install_preview_global.sh +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/kvcache_exactness_audit_20260703.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/kvcache_soak_20260703.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/kvcache_warm_probe_20260703.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/make_fp16_precision_sibling.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/measure_request_capture.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/midform_gate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/model_release_policy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/mtp_cost_probe.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/oc_tap.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/oc_tap_diff.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/opencode_concurrency_qa.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/phase0h_paged_verifier_exactness.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/pillar_gate_qa.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/probe_mlx_pr3026_qsdpa.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/probe_mx_compile_buckets.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/probe_paged_gqa_sdpa_routes.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/qsa_indexer_prefill_numeric_check.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/qsa_prefill_battery2_20260830.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/qsa_prefill_battery_20260830.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/qsa_prefill_flash_numeric_check.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/quality_fp16_parent_logitdiff_20260707.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/qwen35b_mtp_batch_evalplus_guarded.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/qwen35b_mtp_batch_numerics_attribution.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/qwen35b_mtp_batch_numerics_guarded.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/qwen4_quality_component_parity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/qwen4_vision_compiled_parity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/qwen4exp_mtp_tiny_smoke.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/qwen4exp_numeric_check.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/r1_chisquare_verifier_correctness.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/release_macos_v1.sh +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/render_release_notes.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/replay_chat_request.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/restamp_head_quant_packs.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/run_context_degradation_diagnostics.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/run_harness_check.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/select_runtime_wheel.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/serve_openai_mtplx.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/session_cache_followup_qa.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/session_forensics.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/sparkle_rehearsal_kit.sh +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/step_acceptance.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/step_smoke.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/stream_qa_gate.sh +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/train_mtp_adapter_c4.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/upload_head_quant_packs.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/validate_step_mtp_injector.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/walltime-lab/run_project.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/scripts/write_launch_lane_matrix.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/setup.cfg +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_a3b_compiled_target_prefix.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_a3b_mtp_batch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_a3b_mtp_batch_driver.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_a3b_whole_moe.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_adaptive.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_adaptive_dtemp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_admission_keeps_the_conversation.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_agent_prompt_fidelity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_agent_rewrites_off_retries.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_aime_serve_gate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_allow_swap_427.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_answer_memory_stop.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_answer_room_request.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_api_benchmark_contracts.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_app_cli_env_parity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ar_batch_cache_family_guard_420.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ar_batch_penalties.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ar_batch_ragged_ssm_mask.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_aria2_downloader.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_arrays_cache_patch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_artifacts.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_async_decode.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_attention_gate_exactness.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_attention_split.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_background_warmup.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_bank_history_policy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_batch_pump_heartbeat.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_batched_decode.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_batched_greedy_argmax_tiebreak.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_batching_foundation.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_bench_depth_sweep_flags.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_blank_retry_greedy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_block_verify_exact_law.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_bonsai_memory_table.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_bonsai_pack_metadata.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_bracket_tool_rescue.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_build_bonsai_mtplx_pack.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_bundle_native_runtime_wheel.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cache_bank.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cache_state.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_canon_gate_hardening.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_canon_policy_smalls.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_canon_postcommit_extension.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_canonicalize_multimodal.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ccopy_bank_route.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_chat_encode_cache.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_chat_segment_memo.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_chat_turn_segments.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_checkpoint_anchors.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_checkpoint_pre_image.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cli_greedy_draft_coupling.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cli_parity_tools.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_client_controls_default.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_code_eval.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_code_eval_gate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_encode_bounded_units.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_prefix_ram_shadow.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_tier_backlog_admission.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_tier_defaults.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_tier_disk_budget.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_tier_disk_usage_scan.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_tier_eviction_yield.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_tier_foreground_yield.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_tier_incremental_encode.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_tier_min_useful_matched.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_tier_reconcile.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_tier_recurrent_boundary_gate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_tier_stats_cache.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_tier_transactional_replace.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cold_tier_write_budget.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_committed_reasoning_canonicalization.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_committed_token_splice.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_compile_state.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_compiled_ar_wiring.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_compiled_forward.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_compiled_verify_evaluated_weights.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_compiled_verify_tf32_off.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_compressed_tensors.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_config.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_config_profile_precedence.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_constrained.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_context_copy_joined_tokens.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_context_copy_ramp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_context_copy_stats.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_context_degradation_profiles.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_convert_qwen4_exp_fp8_scales.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_copy_block_accept.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_correctors.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_cost_depth_policy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_daemon_client.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_dashboard_endpoints.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_death_signature_macos27.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_decode_growth_guards.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_decoded_state_lifetime.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_adaptive_width_bracket.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_adaptive_width_policy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_attention_island.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_attention_island_bracket.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_attn_proj_wide_m3.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_attn_proj_wide_m3_bracket.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_decode.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_dtypes.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_indexer.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_kernel_paths.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_loader.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_moe_tail.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_moe_tail_bracket.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_mtp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_new_math.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_o_lora.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_parity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_sinkhorn_kernel.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_spec.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_deepseek_v4_swiglu_clamp.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_default_models.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_demotions.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_demotions_generation_sites.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_demotions_graphbank_sites.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_demotions_qsa_prefill_sites.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_dense_decode_max_context.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_dense_mrope.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_dense_mrope_compiled_route.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_dense_mrope_generation.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_dense_vision_compiled_parity_script.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_descriptor_family_collisions.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_descriptors_mimo_tune.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_descriptors_qwen4_exp_tune.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_device_draft_core.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_diagnostics.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_doctor_failed_start.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_download_progress.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_draft_confidence_trace.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_draft_launch_provenance.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_draft_lm_head.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_draft_lm_head_chunked_requantization.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_draft_telemetry_truth.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_draft_temp_policy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_encode_seam_unification.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_endpoint_probe_hygiene.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_engine_session_concurrency.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_engine_session_env.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_env_flag_parsing.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_exact_verify_route.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_exactness_320_319.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_expected_value_greedy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_family_settings.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_fast_sampling.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_final_committed_frontier_byte_skip.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_final_tool_preamble_stream_parity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_finish_stop_origin.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_finished_answers_are_never_regenerated.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_first_token_logprobs_api.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_first_token_logprobs_gemma4.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_fixed_m4_dispatch_guard.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_flash_next_quality_m5_ultra.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_flight_recorder.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_float32_operand.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_forge_cli.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_forge_discover_filter.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_forge_family_serve_depths.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_forge_mixed_convert.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_forge_mtp_key_forms.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_forge_mtp_norm_convention.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_forge_quality_build_script.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_forge_qwen4_exp_lane.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_forge_qwen4_exp_roundtrip.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_forge_verify_depths.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_forkev_telemetry.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_fused_add_rmsnorm_exactness.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gdn_blocked_prefill.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gdn_boundary_retention.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gdn_capture_rows.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gdn_conv_norm_fused.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gdn_conv_norm_rows.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gdn_gated_norm.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gdn_out_fused.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gdn_postconv_fusion.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gdn_postconv_impl_gate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gdn_prefill_prework.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gdn_step_fused.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gdn_tape_headquarter.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gemma4_chunked_prefill.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gemma4_prompt_scoring.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gemma4_rotating_cache_trim.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gemma4_scoring_cache.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gemma4_scoring_policy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gemma4_session_cache.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_gemma4_tokenizer_probe.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_generation_rate_fields.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_generation_store_on_prefill.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_generation_sustained.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_graphbank_compiled_verify.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_graphbank_program_retirement.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_graphbank_trace_lifetime.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_greedy_confidence_runner_telemetry.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_greedy_draft_coupling.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_greedy_trio_ports.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_hc_prefill.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_hc_verify_read.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_health_qwen4_install_reports.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_hermes_native_contract.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_hf_loader.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_hidden_tool_guard_opt_in.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_hy_v3_mtp_backend.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_hy_v3_mtp_graft.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_hygiene_scan.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_hyper_scheduler.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_hyper_v3.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_idle_persistence_pump.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_idle_postcommit_subagent.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_idle_session_release.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_install_macos_script.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_issue_404_macos27_mpp_compat.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_jsonc.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_kernel_selfcheck.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_kernel_selfcheck_bonsai.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_kernel_selfcheck_qwen4_prefill.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_kv_attention_trace_526.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_kv_quant_memory.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_kv_quant_meta_state_rewind_526.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_kv_quant_q4_kernel_route.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_laguna_alt_step.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_laguna_compiled_step.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_laguna_fused.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_laguna_model.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_laguna_steel_attn.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_launch_lane.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_lazy_snapshot_cow.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_lease_keeps_draft_history.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_lease_survives_aborted_prefill.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_lfm2_descriptor_resolution.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_lfm2_fast.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_loop_guard.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_max_idle_watchdog.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_max_lifecycle.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memguard_degraded.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memguard_dense_qsa.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memguard_lease_growth.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memguard_moe_rows.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memguard_ownership.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memguard_pressure_trim.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memguard_queued_persistence.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memguard_release_protection.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memguard_snapshot_width.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memory_limit_g5.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memory_plan.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memory_plan_bonsai.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_memory_pressure_guard.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_metadata_scrub.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_metal_memory_caps.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_midform_gate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mimo_catalog.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mlx_process_env.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_model_catalog.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_model_scheduler.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_model_scheduler_persistence.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_model_updates.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_moe_force_unsorted.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_moe_glu_decode.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_moe_glu_verify.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_moe_packed_projections.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_moe_sorted_gather.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_moe_sorted_gather_nax.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_activation_stats.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_adapters.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_alias_load_path.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_batch_numerics.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_batch_serving.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_depth_n.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_depth_sweep.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_final_state_covers_committed_tokens.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_fused_experts.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_history_cache_only.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_history_settled.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_norm_convention.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_patch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_payload_guards.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_mtp_weightless_degrade.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_nax_attention_hardware.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_nax_detect.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_nax_verify.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ngram_prewarm_option.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ngram_prewarm_reservation.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ngram_row_hotness.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ngram_row_layout.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ngram_table_memory.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_no_mlx_imports.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_non_finite_logits_and_lazy_sampler_law.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_omlx_bridge.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_omlx_bridge_late_system_messages.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_onboarding.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_one_copy_admission.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_one_copy_cancelled_answer.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_one_copy_conversation.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_one_copy_postcommit.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_one_copy_reserve_resize.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_one_copy_restarted_draft_history.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_one_shot_family_contract.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_openai_bridge.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_opencode.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_optimization_profiles.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_orphan_stream_guard_first_line.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_orphan_tool_markup.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_owner_idle_pool.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_pack_contract_states.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_paged_tailmask_elide.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_parallel_tool_prompt_policy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_passive_probe_telemetry.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_penalties.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_penalty_request_wiring.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_persistent_replay.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_phase0h_paged_verifier_exactness.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_phys_footprint_guard.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_policy_fingerprint_stability.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_postcommit_arrival_wait.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_postcommit_cross_session_yield.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_postcommit_prefix_reuse.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_postcommit_resolve_for_request.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_postcommit_tools_plumbing.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_postcommit_wait.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_postcommit_wait_integration.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_postcommit_wait_progress.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_pre_first_token_telemetry.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prefill_admission_shed.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prefill_bench.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prefill_chunk_defaults.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prefill_geometric_tail.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prefill_inforward_boundary_capture.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prefill_reread.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prefill_tps_regression.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_pressure_abort_393.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prism_bonsai_kernels.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prism_hadamard_qwen35.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_profile_default_resolution.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_profiles.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_proj_fusion.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_proj_quant.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_promoted_paged_capacity_526.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prompt_encoding.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prompt_publication_is_request_local.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prompt_scoring_parent.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prompt_scoring_probe.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_prompt_scoring_topk.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_public_cli.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_pull_tls_trust.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_px3_planner_seats.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qmv4_matmul.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_dense_band_sdpa.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_fixed_bank_score_scale.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_flash_skip.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_gate_safety.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_gather.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_gather_rows.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_index_block_writes_evaluated.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_indexer_compile.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_indexer_fused_integration.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_indexer_prefill_mpp_runtime.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_indexer_prefill_runtime_blocks.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_indexer_prefill_simd_topk.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_indexer_prefill_static.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_indexer_prepare_metal.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_indexer_select_metal.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_memory_plan_393.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_mtp_precompute.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_mtp_precompute_wiring.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_pooled_row.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_portable_producer_gate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_prefill_direct.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_prefill_direct_greedy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_prefill_direct_numeric.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_prefill_direct_routing.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_prefill_direct_static.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_prefill_flash_contract.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_prefill_flash_direct_parity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_prefill_flash_static.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_prefill_gather_tier.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_prefill_runtime_gates_static.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_prefill_wiring.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_restore_staging.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_selector_prefix_contract.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_tiled_scores.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qsa_verify_select.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_quant_partials_fp16_526.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen35b_mtp_batch_evalplus_guarded.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen35b_mtp_batch_numerics_attribution.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen35b_mtp_batch_numerics_guarded.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen38_family.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen3_5_mtp_backend.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen3_5_mtp_object_level.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen3_5_mtp_sidecar_norms.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_block_verify.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_bucket_publication.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_capacity_consumers.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_claim_contract.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_compiled_mtp_prepare.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_draft_chain_pipeline.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_draft_device_chain.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_draft_k20_prescatter.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_exp_capture_commit.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_exp_compile.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_exp_gdn_inproj_fusion.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_exp_inforward_boundary_capture.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_exp_qsa_cache.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_exp_qsa_qkv_fusion.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_exp_rope.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_fixed_bank_bf16_parity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_fixed_bank_capacity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_fixed_host_tokens_static.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_fixed_m4_capacity_bucket.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_fixed_m4_copy_windows.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_fixed_m4_float32_admission.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_fixed_m4_gate_cache.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_fixed_m4_growth_layerwise.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_fixed_m4_verify_exactness.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_fixed_m4_vision_admission.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_frspec_native.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_growth_recovery.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_hc_prefill_compile.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_hc_prefill_read_switch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_m4_routed_glu.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_m4_stage3.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_m4_stage3_residual_tail.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_moe_prefill_combine.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_moe_sorted_rows.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_mtp_history_cache_only.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_opdiet.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_ple_first_gather_early.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_ple_gate_exactness.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_ple_lookahead_restored_suffix.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_ple_prefill_lookahead.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_ple_serving_verdicts.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_prefill_dq_gemm.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_prefill_midloop_eval.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_prefill_profile_hc.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_qsa_m4_fused_kv_gather.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_route_kernel.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_rows_gather_scale.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_verify_gate_sites.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_verify_glue.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_verify_sdpa_head_chunk.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_vision_compiled_parity_cpu.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_vision_compiled_parity_script.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_vision_compiled_route_e2e.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_vision_mrope.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_wide_prefill_chunk.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen4_yarn_amplitude.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_qwen_row_owned_router.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ragged_kv_cache.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ram_miss_reason.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_reasoning_stream_split.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_recovery_metrics_attribution.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_release_2114_catalog.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_release_2114_manifest.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_release_notes_template.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_release_pins.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_release_qa_fixwave.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_repetition_stop_legitimate_code.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_repetition_stop_long_cycle.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_request_capture.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_request_log_privacy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_request_observability_golden.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_response_envelope_parity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_retrieval.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_retry_prompt_reuses_served_ids.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_route_tape.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_runtime_ar_final_state.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_runtime_deepseek_v4_o_lora.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_runtime_env_flags_refresh.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_runtime_env_latebind.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_runtime_kpis.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_runtime_model_alias.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_runtime_obs_graphbank.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_runtime_obs_kernel_bails.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_runtime_obs_profiles.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_runtime_obs_stream_holdback.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_runtime_scheduler_histogram.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_runtime_session_sizing.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_sampling.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_scoped_reasoning_history.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_sdpa_gqa_packed.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_sdpa_gqa_packed_quant.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_sdpa_nax_flash.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_sdpa_nax_tile.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_select_runtime_wheel.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_serve_generation_mode_default.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_serve_network_share.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_served_window.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_server_obs_caps_and_health.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_server_obs_postcommit_health.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_server_obs_reasoning_rows.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_server_obs_stream_termination.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_server_openai.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_server_origin_policy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_server_responses.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_session_bank.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_session_bank_boundary_shed.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_session_bank_chain_shed.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_session_bank_env_caps.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_session_bank_gc.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_session_bank_lease_accounting.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_session_bank_protected_terminal.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_session_bank_recurrent_partial_restore.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_session_bank_restore_aliasing.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_session_bank_snapshot_settle.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_session_identity_turn_boundary.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_smart_fan_stall_probe.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_smart_fan_wedged_request.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_snapshot_free_rejection_repair.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_source_wrapper.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_split_partials_fp16_526.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ssd_boundary_repersist.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ssd_spill.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_stable_prefix_boundary.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_start_reuses_chosen_model.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_stats_footer_scope.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_step3p5_mtp_patch.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_stop_resend_session_continuity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_stream_error_receipt.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_stream_guard_env.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_stream_phase_rate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_stream_recovery_paths.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_stream_release_pacer.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_stream_response_tail.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_stream_splitter_contract.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_stream_stall_watchdog.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_stream_transcript_golden.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_stream_visible_cadence.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_sustained_long_context_qa.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_system_memory_guard.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_tail_ar_warm_restore_identity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_tail_config_sampler_presence.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_tail_gemma4_stream_holdback.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_tail_profile_truth.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_tensoroffset_kv_inplace.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_thermal.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_thermal_pressure.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_thermal_sidecar.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_thinking_guard.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_thinking_splitter_marker_quotes.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_tiny_max_tokens_warning_436.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_tokenizer_pretokenizer_repair.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_tool_aware_stream_translator.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_tool_call_hidden_guard_and_key_repair.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_tool_choice_prompt_modes.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_tool_fed_retry_parsed_call.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_tool_nested_args_streaming.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_trace_diagnostics.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_trace_incidents.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_trace_parity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ttft_spans.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_tune_qwen4_exp_verify_strategy.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_tune_record_wizard_parity.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_turboquant.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_turboquant_fallback.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_ui_progress.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_validators.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_verify_crossrow.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_verify_kernels_rehearsal.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_vision_control_literals.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_vision_keyed_partial_restore.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_vision_keyed_restore_rules.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_vision_restore_span_guard.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_vision_session_cache.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_vision_session_frontier_gate.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_vision_session_restore.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_vision_tower.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_vision_trunk_forward_positions.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_vllm_reference.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_wide_prefill_chunk_after_reclaim.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/tests/test_working_set.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/README.md +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/__init__.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/build.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/constants.py +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v1/copy_blocks.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v1/float8.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v1/gather_kv_cache.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v1/kv_scale_update.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v1/pagedattention.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v1/reshape_and_cache.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v1/utils.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v2/copy_blocks.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v2/float8.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v2/gather_kv_cache.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v2/gdn_linear_attention.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v2/kv_scale_update.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v2/pagedattention.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v2/reshape_and_cache.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v2/turboquant.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/kernels_v2/utils.metal +0 -0
- {mtplx-2.12.1 → mtplx-2.12.2}/vllm_metal/metal/paged_ops.cpp +0 -0
|
@@ -4,6 +4,20 @@ All notable user-facing changes to MTPLX. The format is based on
|
|
|
4
4
|
[Keep a Changelog](https://keepachangelog.com/en/1.1.0/); versions follow
|
|
5
5
|
[Semantic Versioning](https://semver.org/).
|
|
6
6
|
|
|
7
|
+
## [2.12.2] - 2026-10-03
|
|
8
|
+
|
|
9
|
+
### Fixed
|
|
10
|
+
|
|
11
|
+
- **Macs with 8, 16 and 32 GB serve ordinary prompts again.** 2.12.1 refused prompts that 2.12.0 served on these Macs with the model the app picks for them. On an M5 Max set up as each Mac (its engine limit, its free-memory floors and a fixed free-memory reading), 2.12.1 refused an 18-token prompt to Qwen 3.5 4B on 8 GB with 3 GiB free, 2,600 to 7,800-token prompts and an agent's first turn to Ternary Bonsai 2 27B on 16 GB, and 6,800 and 7,800-token prompts and an agent's third turn to Qwen 3.8 27B Optimized Speed on 32 GB; GitHub's 7 GB M1 test machine refused a one-line request with a 507. 2.12.2 serves all of them (October 3), and the M1 test machine passes. Three changes:
|
|
12
|
+
- A dense model's prompt is priced at what it was measured to use through `mtplx serve`: the 4B 0.19 GiB for a 59-token prompt and 1.55 to 1.65 GiB per 2,048-token chunk, Bonsai 2 27B 0.45 GiB for a 98-token prompt, Bonsai and the 27B 2.5 to 2.7 GiB per 2,048-token chunk, plus a margin of 0.09 to 0.25 GiB. 2.12.1 charged every model the 27B's 3 GiB per 2,048-token chunk and about 2 GiB for a one-line prompt. Models with routed experts, Flash-Next and Gemma 4 keep their 2.12.1 prices.
|
|
13
|
+
- A prompt that does not fit in 2,048-token chunks runs in 1,024-token chunks, or 512-token chunks for prompts up to about 8,000 tokens, before it is refused (not Flash-Next or Gemma 4, which keep their own chunk widths). A narrower chunk is used only where it saves at least 128 MiB. A 1,024-token chunk uses about 1 GiB less on the 27B models; a 7,000-token Bonsai prompt took 9.5 s at 2,048, 9.6 s at 1,024 and 9.7 s at 512 tokens a chunk (single runs).
|
|
14
|
+
- Memory outside the GPU allocator is charged only past 4 GiB on Macs under 64 GB. 2.12.1 charged it past a sixteenth of RAM (1 GiB on 16 GB, 2 GiB on 32 GB), while a healthy engine holds 2.1 to 2.9 GiB there with Bonsai and 2.3 to 4.0 GiB with the 27B. Macs with 64 GB or more are unchanged.
|
|
15
|
+
- **#499's 99,355-token 27B turn on a 48 GB Mac runs in 1,024-token chunks**, priced at 34.87 GiB against the 36 GiB limit, where 2.12.1 refused it (tests with the reporter's numbers).
|
|
16
|
+
|
|
17
|
+
### Changed
|
|
18
|
+
|
|
19
|
+
- **Dependencies.** urllib3 2.8.0 (#582 by @dependabot).
|
|
20
|
+
|
|
7
21
|
## [2.12.1] - 2026-10-02
|
|
8
22
|
|
|
9
23
|
### Added
|
|
@@ -20891,6 +20891,15 @@ def _prefill_forward_bill(
|
|
|
20891
20891
|
scratch, source = _admission_scratch_bytes(
|
|
20892
20892
|
state, rows=rows, prompt_tokens=max(1, int(prompt_tokens)), geometry=geometry
|
|
20893
20893
|
)
|
|
20894
|
+
if (
|
|
20895
|
+
source.startswith("geometry_measured")
|
|
20896
|
+
and width is not None
|
|
20897
|
+
and int(width) < _DENSE_PREFILL_FLAT_WIDTH
|
|
20898
|
+
):
|
|
20899
|
+
# Under 1,024 rows a dense family's chunks also hold about a live
|
|
20900
|
+
# row per token of context (_ADMISSION_NARROW_PREFILL_WIDTHS).
|
|
20901
|
+
scratch += max(1, int(prompt_tokens)) * int(geometry.live_bytes_per_token)
|
|
20902
|
+
source += "+context"
|
|
20894
20903
|
chunk = _admission_chunk_bytes(geometry, rows, scratch)
|
|
20895
20904
|
return {"rows": rows, "scratch": scratch, "source": source, "chunk_bytes": chunk}
|
|
20896
20905
|
|
|
@@ -21135,10 +21144,15 @@ def _vision_bank_session_id(bank: Any, prompt_ids: list[int], splice: Any) -> st
|
|
|
21135
21144
|
# allowance is that measurement plus headroom on the large seats, and a
|
|
21136
21145
|
# sixteenth of the machine below 128 GB: 8 GiB outside Metal on a 16 GB Mac
|
|
21137
21146
|
# (limit 12 GiB) would put the guard's process ceiling past the RAM itself.
|
|
21138
|
-
#
|
|
21147
|
+
# It never goes under what a healthy daemon of a small seat's own model
|
|
21148
|
+
# holds: 2.1 to 2.9 GiB for Bonsai 2 27B and 2.3 to 4.0 GiB for the 27B
|
|
21149
|
+
# (2026-10-02, at the 16 and 32 GB limits; the 4.0 right after a prefill
|
|
21150
|
+
# chunk, its freed buffers not yet given back). 2.12.1's sixteenth (1 GiB on
|
|
21151
|
+
# 16 GB, 2 on 32) charged that as a leak and refused prompts 2.12.0 served.
|
|
21152
|
+
# 8 to 64 GB: 4 GiB, 96 GB: 6 GiB, 128 GB and up: 8 GiB.
|
|
21139
21153
|
# MTPLX_HOST_MEMORY_ALLOWANCE_BYTES overrides it.
|
|
21140
21154
|
_HOST_MEMORY_ALLOWANCE_CAP_BYTES = 8 * 1024**3
|
|
21141
|
-
_HOST_MEMORY_ALLOWANCE_MIN_BYTES =
|
|
21155
|
+
_HOST_MEMORY_ALLOWANCE_MIN_BYTES = 4 * 1024**3
|
|
21142
21156
|
_HOST_MEMORY_ALLOWANCE_RAM_DIVISOR = 16
|
|
21143
21157
|
|
|
21144
21158
|
|
|
@@ -21471,6 +21485,80 @@ _ADMISSION_LIVE_LAYERS = 4
|
|
|
21471
21485
|
# Flash-Next (generation._WIDE_PREFILL_FIXED_BYTES).
|
|
21472
21486
|
_ADMISSION_FIXED_FLOOR_BYTES = 1 * 1024**3
|
|
21473
21487
|
|
|
21488
|
+
# A dense family's prefill working memory (no routed experts, no QSA
|
|
21489
|
+
# indexer: the 27B, Bonsai 2 27B, the 4B and 9B), measured 2026-10-02 on an
|
|
21490
|
+
# M5 Max through ``mtplx serve``: MLX's peak over what a chunk leaves
|
|
21491
|
+
# resident, with the peak reset before each fresh prompt, in GiB.
|
|
21492
|
+
#
|
|
21493
|
+
# rows 59-195 308-345 606-641 1,024 2,048
|
|
21494
|
+
# Qwen 3.5 4B 0.19-0.38 0.69 0.93 1.06-1.24 1.55-1.65
|
|
21495
|
+
# Bonsai 2 27B 0.45-0.74 0.93 1.30 1.70-1.72 2.69-2.72
|
|
21496
|
+
# Qwen 3.8 27B 1.52-1.54 2.52-2.55
|
|
21497
|
+
#
|
|
21498
|
+
# At 1,024 and 2,048 rows it stays flat from 2K to 49K tokens of context. It
|
|
21499
|
+
# climbs steeply over the first couple of hundred rows and then by about ten
|
|
21500
|
+
# MLP rows (2 x (hidden + 3 x intermediate)) a row, so the bill is the lower
|
|
21501
|
+
# of two lines in MLP rows: 0.25 GiB plus 36 a row, and 0.75 GiB plus 10 a
|
|
21502
|
+
# row. That covers every reading by 0.09 to 0.25 GiB: 0.37 GiB for a
|
|
21503
|
+
# 59-token 4B prompt, 1.84 and 2.94 GiB at 1,024 and 2,048 rows on the 27B
|
|
21504
|
+
# geometry, 1.33 and 1.90 on the 4B. 2.12.1 charged every family the 27B's
|
|
21505
|
+
# whole-request reserve per 2,048 rows with a 1.9 to 2.2 GiB floor (3.0 GiB
|
|
21506
|
+
# at 2,048 rows, 2.3 for a 64-token 4B prompt), which refused short prompts
|
|
21507
|
+
# on 8 GB Macs. A family with routed experts keeps that bill until it is
|
|
21508
|
+
# measured.
|
|
21509
|
+
_DENSE_PREFILL_SHORT_FIXED_BYTES = 256 * 1024**2
|
|
21510
|
+
_DENSE_PREFILL_SHORT_MLP_ROWS = 36
|
|
21511
|
+
_DENSE_PREFILL_FIXED_BYTES = 768 * 1024**2
|
|
21512
|
+
_DENSE_PREFILL_MLP_ROWS = 10
|
|
21513
|
+
|
|
21514
|
+
# The chunk widths the admission may narrow a prefill to before it refuses
|
|
21515
|
+
# (``_admission_narrow_widths``). From 1,024 rows up a dense family's chunk
|
|
21516
|
+
# does not grow with the context. At 512 rows it holds about one more live
|
|
21517
|
+
# row per token of context (the 4B: 0.94 GiB at 7K, 1.68 at 32K, 2.22 at
|
|
21518
|
+
# 49K; Bonsai 2 27B: 1.21 at 2.4K, 1.53 at 8K), which
|
|
21519
|
+
# ``_prefill_forward_bill`` adds, so a 512-row chunk costs less than a
|
|
21520
|
+
# 1,024-row one only for short and medium prompts. The admission keeps a
|
|
21521
|
+
# narrower width only where it costs less than the width above it. A
|
|
21522
|
+
# 7K-token Bonsai prompt took 9.5 s at 2,048 rows, 9.6 s at 1,024 and 9.7 s
|
|
21523
|
+
# at 512; a 16K-token 4B prompt 3.4, 3.6 and 4.1 s.
|
|
21524
|
+
_ADMISSION_NARROW_PREFILL_WIDTHS = (1024, 512)
|
|
21525
|
+
_DENSE_PREFILL_FLAT_WIDTH = 1024
|
|
21526
|
+
# A narrower chunk runs slower; one that saves less than this over a wider
|
|
21527
|
+
# one is not run in its place when nothing fits.
|
|
21528
|
+
_NARROW_CHUNK_MIN_SAVING_BYTES = 128 * 1024**2
|
|
21529
|
+
|
|
21530
|
+
|
|
21531
|
+
def _dense_prefill_bill(args: Any, rows: int) -> int | None:
|
|
21532
|
+
"""What a prefill forward of ``rows`` rows of a dense family holds at its
|
|
21533
|
+
peak (``_DENSE_PREFILL_FIXED_BYTES``). None for a config without a dense
|
|
21534
|
+
MLP, or with routed experts."""
|
|
21535
|
+
|
|
21536
|
+
def value(name: str) -> int:
|
|
21537
|
+
try:
|
|
21538
|
+
return int(getattr(args, name, 0) or 0)
|
|
21539
|
+
except (TypeError, ValueError):
|
|
21540
|
+
return 0
|
|
21541
|
+
|
|
21542
|
+
hidden = value("hidden_size")
|
|
21543
|
+
intermediate = value("intermediate_size")
|
|
21544
|
+
if hidden <= 0 or intermediate <= 0:
|
|
21545
|
+
return None
|
|
21546
|
+
if (
|
|
21547
|
+
value("num_experts_per_tok")
|
|
21548
|
+
or value("top_k_experts")
|
|
21549
|
+
or value("moe_topk")
|
|
21550
|
+
or value("num_experts")
|
|
21551
|
+
or value("n_routed_experts")
|
|
21552
|
+
or value("num_local_experts")
|
|
21553
|
+
):
|
|
21554
|
+
return None
|
|
21555
|
+
mlp_row = 2 * (hidden + 3 * intermediate)
|
|
21556
|
+
rows = max(1, int(rows))
|
|
21557
|
+
return min(
|
|
21558
|
+
_DENSE_PREFILL_SHORT_FIXED_BYTES + _DENSE_PREFILL_SHORT_MLP_ROWS * mlp_row * rows,
|
|
21559
|
+
_DENSE_PREFILL_FIXED_BYTES + _DENSE_PREFILL_MLP_ROWS * mlp_row * rows,
|
|
21560
|
+
)
|
|
21561
|
+
|
|
21474
21562
|
|
|
21475
21563
|
def _forward_row_bytes(args: Any) -> int | None:
|
|
21476
21564
|
"""Bytes one row holds at the peak of one layer's forward, for the
|
|
@@ -21574,6 +21662,10 @@ def _admission_scratch_bytes(
|
|
|
21574
21662
|
materializes below the sparse crossover, and a fixed part; about 3.5 GiB
|
|
21575
21663
|
at 2,048 rows for a 74K prompt.
|
|
21576
21664
|
|
|
21665
|
+
A dense family (no routed experts) is charged what was measured on it
|
|
21666
|
+
(``_DENSE_PREFILL_FIXED_BYTES``): 1.84 and 2.94 GiB at 1,024 and 2,048
|
|
21667
|
+
rows on the 27B, 1.33 and 1.90 on the 4B, 0.37 GiB for a 59-token prompt.
|
|
21668
|
+
|
|
21577
21669
|
Every other family is charged from its geometry: the rows the forward
|
|
21578
21670
|
runs times what a row holds across the live layers
|
|
21579
21671
|
(``_forward_row_bytes``), plus a fixed part. The one measurement on the
|
|
@@ -21632,6 +21724,10 @@ def _admission_scratch_bytes(
|
|
|
21632
21724
|
# describe it; its activations, the fixed part and its measured
|
|
21633
21725
|
# attention do.
|
|
21634
21726
|
return fixed + per_row * rows + attention, "geometry_calibrated" + suffix
|
|
21727
|
+
dense = _dense_prefill_bill(_runtime_text_args(runtime), rows)
|
|
21728
|
+
if dense:
|
|
21729
|
+
# Measured on this family (``_DENSE_PREFILL_FIXED_BYTES``).
|
|
21730
|
+
return dense + attention, "geometry_measured" + suffix
|
|
21635
21731
|
return max(fixed + per_row * rows, flat_share) + attention, "geometry" + suffix
|
|
21636
21732
|
|
|
21637
21733
|
|
|
@@ -22740,6 +22836,31 @@ def _admission_prefill_widths(
|
|
|
22740
22836
|
return prefill_forward_widths(runtime, prompt_tokens, requested)
|
|
22741
22837
|
|
|
22742
22838
|
|
|
22839
|
+
def _admission_narrow_widths(runtime: Any, widths: Sequence[int | None]) -> list[int]:
|
|
22840
|
+
"""The narrower chunks the admission may run a family this module chunks
|
|
22841
|
+
(the 27B, Bonsai, the 4B and 9B) at, below its ladder
|
|
22842
|
+
(``_ADMISSION_NARROW_PREFILL_WIDTHS``): a chunk's working memory grows
|
|
22843
|
+
with its rows, so a Mac that cannot fit the profile's chunk runs a
|
|
22844
|
+
narrower one instead of being refused (2.12.1 refused ordinary prompts on
|
|
22845
|
+
16 and 32 GB Macs that a 1,024-row chunk fits). None for Flash-Next,
|
|
22846
|
+
which keeps its own ladder, a backend that answers for its own widths,
|
|
22847
|
+
or a prompt forwarded whole; 1,024 only for a family with routed
|
|
22848
|
+
experts, whose 512-row chunks were not measured."""
|
|
22849
|
+
|
|
22850
|
+
if (
|
|
22851
|
+
runtime is None
|
|
22852
|
+
or not widths
|
|
22853
|
+
or widths[-1] is None
|
|
22854
|
+
or callable(getattr(runtime, "prefill_forward_widths", None))
|
|
22855
|
+
or _runtime_has_qsa_indexer(runtime)
|
|
22856
|
+
):
|
|
22857
|
+
return []
|
|
22858
|
+
rungs = _ADMISSION_NARROW_PREFILL_WIDTHS
|
|
22859
|
+
if _dense_prefill_bill(_runtime_text_args(runtime), 1) is None:
|
|
22860
|
+
rungs = tuple(w for w in rungs if w >= _DENSE_PREFILL_FLAT_WIDTH)
|
|
22861
|
+
return [w for w in rungs if w < int(widths[-1])]
|
|
22862
|
+
|
|
22863
|
+
|
|
22743
22864
|
def _run_prefill_admission(
|
|
22744
22865
|
state: "ServerState",
|
|
22745
22866
|
*,
|
|
@@ -22842,6 +22963,11 @@ def _run_prefill_admission(
|
|
|
22842
22963
|
# at 2,048 rows because the wide gate refused before a reclamation
|
|
22843
22964
|
# that left room for 4,096).
|
|
22844
22965
|
widths = sorted({*(int(rung) for rung in wide_prefill_rungs), *widths}, reverse=True)
|
|
22966
|
+
# The narrower chunks a tight Mac may fall back to (the 27B, Bonsai, the
|
|
22967
|
+
# 4B and 9B); ``cheapest`` below keeps them from running slower for
|
|
22968
|
+
# nothing.
|
|
22969
|
+
narrow_rungs = [] if prompt_scoring else _admission_narrow_widths(runtime, widths)
|
|
22970
|
+
widths = list(widths) + narrow_rungs
|
|
22845
22971
|
output_tokens = int(
|
|
22846
22972
|
_dynamic_paged_kv_initial_new_token_budget(max_new_tokens)[0]
|
|
22847
22973
|
) + max(0, int(mtp_depth or 0))
|
|
@@ -23119,6 +23245,22 @@ def _run_prefill_admission(
|
|
|
23119
23245
|
return w
|
|
23120
23246
|
return _ADMISSION_NO_FIT
|
|
23121
23247
|
|
|
23248
|
+
def cheapest(models: Mapping[Any, Any]) -> Any:
|
|
23249
|
+
# The width reclamation prices and a request that fits nowhere runs
|
|
23250
|
+
# at. With the narrower rungs, the widest one within
|
|
23251
|
+
# _NARROW_CHUNK_MIN_SAVING_BYTES of the least growth: when the
|
|
23252
|
+
# request's peak is its repage or the start of its answer, or a
|
|
23253
|
+
# 512-row chunk's context outgrows the 1,024-row one, a narrower
|
|
23254
|
+
# chunk costs no less and only runs slower. Without them, the
|
|
23255
|
+
# ladder's last width, as before.
|
|
23256
|
+
if not narrow_rungs:
|
|
23257
|
+
return widths[-1]
|
|
23258
|
+
least = min(int(models[w]["growth_bytes"]) for w in widths)
|
|
23259
|
+
for w in widths:
|
|
23260
|
+
if int(models[w]["growth_bytes"]) <= least + _NARROW_CHUNK_MIN_SAVING_BYTES:
|
|
23261
|
+
return w
|
|
23262
|
+
return widths[-1]
|
|
23263
|
+
|
|
23122
23264
|
models = price()
|
|
23123
23265
|
chosen = widest_fit(now, models)
|
|
23124
23266
|
early_pool_clear: dict[str, Any] | None = None
|
|
@@ -23203,7 +23345,7 @@ def _run_prefill_admission(
|
|
|
23203
23345
|
),
|
|
23204
23346
|
}
|
|
23205
23347
|
return None
|
|
23206
|
-
narrow =
|
|
23348
|
+
narrow = cheapest(models)
|
|
23207
23349
|
current = models[narrow if chosen is _ADMISSION_NO_FIT else chosen]
|
|
23208
23350
|
receipt: dict[str, Any] = {
|
|
23209
23351
|
"action": "prefill_admission_shed",
|
|
@@ -23310,12 +23452,13 @@ def _run_prefill_admission(
|
|
|
23310
23452
|
def replan(step: str, evicted: Any) -> None:
|
|
23311
23453
|
# A step that evicted anything may have changed what the restore
|
|
23312
23454
|
# reads: ask again and price the request on the answer.
|
|
23313
|
-
nonlocal restore, models, current
|
|
23455
|
+
nonlocal restore, models, current, narrow
|
|
23314
23456
|
if not evicted:
|
|
23315
23457
|
return
|
|
23316
23458
|
before = restore
|
|
23317
23459
|
restore = plan_restore()
|
|
23318
23460
|
models = price()
|
|
23461
|
+
narrow = cheapest(models)
|
|
23319
23462
|
current = models[narrow]
|
|
23320
23463
|
if (
|
|
23321
23464
|
restore["reused_tokens"] != before["reused_tokens"]
|
|
@@ -23387,8 +23530,9 @@ def _run_prefill_admission(
|
|
|
23387
23530
|
if not one_copy and deficit(now) > 0 and current["publish_copy_bytes"] > 0:
|
|
23388
23531
|
flags["skip_publish"] = True
|
|
23389
23532
|
unpublished = price()
|
|
23390
|
-
if unpublished[
|
|
23533
|
+
if unpublished[cheapest(unpublished)]["growth_bytes"] < current["growth_bytes"]:
|
|
23391
23534
|
models = unpublished
|
|
23535
|
+
narrow = cheapest(models)
|
|
23392
23536
|
current = models[narrow]
|
|
23393
23537
|
receipt["prompt_publish_skipped"] = True
|
|
23394
23538
|
steps.append("prompt_publish")
|
|
@@ -959,6 +959,7 @@ tests/test_session_bank_recurrent_partial_restore.py
|
|
|
959
959
|
tests/test_session_bank_restore_aliasing.py
|
|
960
960
|
tests/test_session_bank_snapshot_settle.py
|
|
961
961
|
tests/test_session_identity_turn_boundary.py
|
|
962
|
+
tests/test_small_mac_prefill.py
|
|
962
963
|
tests/test_smart_fan_stall_probe.py
|
|
963
964
|
tests/test_smart_fan_wedged_request.py
|
|
964
965
|
tests/test_snapshot_free_rejection_repair.py
|
|
@@ -23,6 +23,12 @@ MTPLX_MEMORY_LIMIT_BYTES did not lower it, and 14 GiB of leaked host memory
|
|
|
23
23
|
ceiling (12 GiB limit + 8) past its RAM. The seats that motivated the old
|
|
24
24
|
rule stay quiet: 48 GB / 27B with 3 GiB of host memory, 96 GB / Flash-Next
|
|
25
25
|
with 5.
|
|
26
|
+
|
|
27
|
+
2026-10-02: never under 4 GiB. A sixteenth of a small Mac is less than what
|
|
28
|
+
a healthy daemon of that Mac's own model holds outside MLX (2.1 to 2.9 GiB
|
|
29
|
+
for Bonsai 2 27B at the 16 GB limit, 2.3 to 4.0 GiB for the 27B at the 32
|
|
30
|
+
GB limit), so 2.12.1 charged ordinary memory as a leak and refused prompts
|
|
31
|
+
2.12.0 served. 8 to 64 GB: 4 GiB, 96 GB: 6, 128 GB and up: 8.
|
|
26
32
|
"""
|
|
27
33
|
|
|
28
34
|
from __future__ import annotations
|
|
@@ -89,13 +95,15 @@ def _served_prefill_is_chunked(monkeypatch):
|
|
|
89
95
|
|
|
90
96
|
@pytest.mark.parametrize(
|
|
91
97
|
"ram_gib, allowance_gib",
|
|
92
|
-
[(16,
|
|
98
|
+
[(8, 4), (16, 4), (32, 4), (48, 4), (64, 4), (96, 6), (128, 8), (512, 8)],
|
|
93
99
|
)
|
|
94
100
|
def test_the_allowance_is_sized_to_the_seat(ram_gib, allowance_gib):
|
|
95
|
-
"""A sixteenth of the machine
|
|
101
|
+
"""A sixteenth of the machine between 4 GiB and the 8 GiB measured on
|
|
102
|
+
128 GB.
|
|
96
103
|
|
|
97
104
|
A flat 8 GiB put a 16 GB Mac's process ceiling (12 GiB limit + 8) past
|
|
98
|
-
its RAM; the 2026-09-27 review of this change caught it.
|
|
105
|
+
its RAM; the 2026-09-27 review of this change caught it. A sixteenth
|
|
106
|
+
alone (1 GiB on 16 GB) charged a healthy daemon's own host memory."""
|
|
99
107
|
|
|
100
108
|
state = _state(total_gib=ram_gib, limit_gib=ram_gib * 0.75)
|
|
101
109
|
assert srv._host_memory_allowance_bytes(state) == allowance_gib * GIB
|
|
@@ -103,8 +111,10 @@ def test_the_allowance_is_sized_to_the_seat(ram_gib, allowance_gib):
|
|
|
103
111
|
|
|
104
112
|
def test_a_simulated_seat_gets_its_own_allowance():
|
|
105
113
|
# --memory-budget 48G on a 128 GB Mac plans (and guards) a 48 GB seat.
|
|
114
|
+
state = _state(total_gib=128, limit_gib=36, budget_gib=96)
|
|
115
|
+
assert srv._host_memory_allowance_bytes(state) == 6 * GIB
|
|
106
116
|
state = _state(total_gib=128, limit_gib=36, budget_gib=48)
|
|
107
|
-
assert srv._host_memory_allowance_bytes(state) ==
|
|
117
|
+
assert srv._host_memory_allowance_bytes(state) == 4 * GIB
|
|
108
118
|
|
|
109
119
|
|
|
110
120
|
def test_an_unknown_machine_gets_the_cap():
|
|
@@ -241,12 +251,12 @@ def test_admission_does_not_refuse_a_request_the_plan_sized_to_fit(monkeypatch):
|
|
|
241
251
|
|
|
242
252
|
|
|
243
253
|
def test_a_leak_on_a_48gb_seat_is_charged_not_forgiven(monkeypatch):
|
|
244
|
-
"""
|
|
245
|
-
allowance (8 GiB on this seat) charged 2 GiB. The seat's
|
|
254
|
+
"""11 GiB outside MLX on a 48 GB Mac. This test used to admit 10: the old
|
|
255
|
+
allowance (8 GiB on this seat) charged 2 GiB. The seat's 4 GiB allowance
|
|
246
256
|
charges 7, and the same request no longer fits the 36 GiB limit."""
|
|
247
257
|
|
|
248
258
|
state = _state(total_gib=48, limit_gib=36)
|
|
249
|
-
_pin(monkeypatch, allocator_gib=28, footprint_gib=
|
|
259
|
+
_pin(monkeypatch, allocator_gib=28, footprint_gib=39)
|
|
250
260
|
receipt = srv._prefill_admission_shed(
|
|
251
261
|
state, prompt_ids=list(range(8192)), session_bank=_EmptyBank(), session_id="pi"
|
|
252
262
|
)
|
|
@@ -275,3 +285,48 @@ def test_the_admission_receipt_explains_what_it_charged(monkeypatch):
|
|
|
275
285
|
# Still over the limit after reclamation, on the footprint alone.
|
|
276
286
|
assert receipt["refused"] is True
|
|
277
287
|
assert receipt["host_overhang_charged_bytes_after"] == 32 * GIB
|
|
288
|
+
|
|
289
|
+
|
|
290
|
+
# --------------------------------------------------------------------------
|
|
291
|
+
# Small Macs (2.12.2)
|
|
292
|
+
# --------------------------------------------------------------------------
|
|
293
|
+
|
|
294
|
+
|
|
295
|
+
MIB = 1024**2
|
|
296
|
+
|
|
297
|
+
|
|
298
|
+
@pytest.mark.parametrize(
|
|
299
|
+
"ram_gib, limit_gib, allocator_mib, footprint_mib, charged_mib",
|
|
300
|
+
[
|
|
301
|
+
# Bonsai 2 27B at the 16 GB limit: 2.9 GiB outside MLX after a prompt
|
|
302
|
+
# (2.12.1 charged 1.9 GiB of it).
|
|
303
|
+
(16, 12, 7782, 10772, 0),
|
|
304
|
+
# The 27B at the 32 GB limit before its prefill: 2.3 GiB.
|
|
305
|
+
(32, 24, 20470, 22820, 0),
|
|
306
|
+
# The same daemon right after a prefill chunk, its freed buffers not
|
|
307
|
+
# yet given back: 4.015 GiB, of which 2.12.1 charged 2.015.
|
|
308
|
+
(32, 24, 21064, 25175, 15),
|
|
309
|
+
],
|
|
310
|
+
)
|
|
311
|
+
def test_a_small_macs_own_model_is_not_charged(
|
|
312
|
+
monkeypatch, ram_gib, limit_gib, allocator_mib, footprint_mib, charged_mib
|
|
313
|
+
):
|
|
314
|
+
"""The readings 2.12.1 refused on (2026-10-02, at the 16 and 32 GB limits)."""
|
|
315
|
+
|
|
316
|
+
state = _state(total_gib=ram_gib, limit_gib=limit_gib)
|
|
317
|
+
_pin(monkeypatch, allocator_gib=allocator_mib / 1024, footprint_gib=footprint_mib / 1024)
|
|
318
|
+
live, fields = srv._footprint_floor(
|
|
319
|
+
state, limit=int(limit_gib * GIB), allocator_bytes=allocator_mib * MIB
|
|
320
|
+
)
|
|
321
|
+
assert fields["host_allowance_bytes"] == 4 * GIB
|
|
322
|
+
assert fields["host_overhang_charged_bytes"] == charged_mib * MIB
|
|
323
|
+
assert live == (allocator_mib + charged_mib) * MIB
|
|
324
|
+
|
|
325
|
+
|
|
326
|
+
def test_a_leak_on_a_16gb_seat_is_still_charged(monkeypatch):
|
|
327
|
+
# 6 GiB outside MLX on a 16 GB Mac: 2 GiB past its 4 GiB allowance.
|
|
328
|
+
state = _state(total_gib=16, limit_gib=12)
|
|
329
|
+
_pin(monkeypatch, allocator_gib=9, footprint_gib=15)
|
|
330
|
+
_live, fields = srv._footprint_floor(state, limit=12 * GIB, allocator_bytes=9 * GIB)
|
|
331
|
+
assert fields["host_allowance_bytes"] == 4 * GIB
|
|
332
|
+
assert fields["host_overhang_charged_bytes"] == 2 * GIB
|
|
@@ -574,13 +574,15 @@ class Test499FortyEightGigSeat:
|
|
|
574
574
|
assert receipt.get("refused") is not True
|
|
575
575
|
assert receipt["projected_bytes_after"] <= 36 * GIB
|
|
576
576
|
|
|
577
|
-
def
|
|
577
|
+
def test_with_the_heads_history_the_turn_runs_at_a_narrower_chunk(self, monkeypatch):
|
|
578
578
|
"""The plan now carries the MTP head's committed history (4,096 B a
|
|
579
579
|
token on the 27B). Priced with it the turn projects 39.47 GiB, the
|
|
580
|
-
trace's measured 39.5 GiB peak,
|
|
581
|
-
|
|
582
|
-
|
|
583
|
-
|
|
580
|
+
trace's measured 39.5 GiB peak, where 2.12.0 ran it past the limit and
|
|
581
|
+
produced no tokens before the client gave up. Without the banked
|
|
582
|
+
prompt copy it needed 36.03 GiB at 2,048 rows and was refused (2.12.1);
|
|
583
|
+
at 1,024 rows, whose measured working memory is 1.1 GiB less
|
|
584
|
+
(2026-10-02: 1.52 against 2.55 GiB on the 27B), it needs 34.87 GiB,
|
|
585
|
+
under the 0.97 line of the 36 GiB limit, and runs."""
|
|
584
586
|
|
|
585
587
|
plan = self._plan()
|
|
586
588
|
plan.mtp_history_bytes_per_token = 4_096
|
|
@@ -609,13 +611,10 @@ class Test499FortyEightGigSeat:
|
|
|
609
611
|
session.end_generation()
|
|
610
612
|
assert 39.4 * GIB < receipt["projected_bytes"] < 39.5 * GIB
|
|
611
613
|
assert receipt["prompt_publish_skipped"] is True
|
|
612
|
-
assert
|
|
613
|
-
assert receipt["
|
|
614
|
-
assert
|
|
615
|
-
|
|
616
|
-
# turn needs: a shorter prompt, not a retry, fits.
|
|
617
|
-
assert receipt["retry_can_succeed"] is False
|
|
618
|
-
assert receipt["retry_when"] == "not_without_a_shorter_prompt"
|
|
614
|
+
assert "refused" not in receipt
|
|
615
|
+
assert receipt["prefill_chunk_requested"] == 2048
|
|
616
|
+
assert receipt["prefill_chunk_tokens"] == 1024
|
|
617
|
+
assert 34.8 * GIB < receipt["projected_bytes_after"] < 0.97 * 36 * GIB
|
|
619
618
|
|
|
620
619
|
def test_the_banked_prompt_copy_is_priced_while_it_fits_the_cap(self, monkeypatch):
|
|
621
620
|
plan = self._plan()
|
|
@@ -705,15 +704,15 @@ class TestGrowthModel:
|
|
|
705
704
|
def test_the_27b_is_priced_from_its_geometry(self):
|
|
706
705
|
"""#525 point A: a family without a QSA indexer is billed rows times
|
|
707
706
|
what a row holds plus a fixed part, never 0 and never a flat figure
|
|
708
|
-
that ignores the rows. The 27B
|
|
709
|
-
|
|
710
|
-
|
|
711
|
-
|
|
712
|
-
rows."""
|
|
707
|
+
that ignores the rows. The 27B is a dense family, billed what was
|
|
708
|
+
measured on it (2026-10-02, the 27B and Bonsai 2 27B through mtplx
|
|
709
|
+
serve: 0.45 to 0.74 GiB for 98 to 195 rows, 1.52 to 1.72 at 1,024,
|
|
710
|
+
2.52 to 2.72 at 2,048): 0.75 GiB plus ten MLP rows a row (1.09 MiB).
|
|
711
|
+
2.12.1 charged 3 GiB at 2,048 rows and 2.04 GiB for a 195-token turn."""
|
|
713
712
|
|
|
714
713
|
args = _q27_text_args()
|
|
715
|
-
per_row = srv._ADMISSION_LIVE_LAYERS * srv._forward_row_bytes(args)
|
|
716
714
|
assert srv._forward_row_bytes(args) == 139_264
|
|
715
|
+
mlp_row = 2 * (5_120 + 3 * 17_408)
|
|
717
716
|
state = SimpleNamespace(runtime=_q27_runtime())
|
|
718
717
|
geometry = self._geometry(live_bytes_per_token=Q27_KV)
|
|
719
718
|
|
|
@@ -721,14 +720,17 @@ class TestGrowthModel:
|
|
|
721
720
|
scratch, source = srv._admission_scratch_bytes(
|
|
722
721
|
state, rows=rows, prompt_tokens=prompt_tokens, geometry=geometry
|
|
723
722
|
)
|
|
724
|
-
assert source == "
|
|
723
|
+
assert source == "geometry_measured"
|
|
725
724
|
return scratch
|
|
726
725
|
|
|
727
|
-
|
|
728
|
-
assert bill(2048)
|
|
729
|
-
assert bill(
|
|
730
|
-
|
|
731
|
-
assert bill(
|
|
726
|
+
assert bill(2048) == 0.75 * GIB + 10 * mlp_row * 2048
|
|
727
|
+
assert 2.72 * GIB < bill(2048) < 2.95 * GIB
|
|
728
|
+
assert 1.72 * GIB < bill(1024) < 1.85 * GIB
|
|
729
|
+
# Short prompts are on the steeper line: 0.25 GiB plus 36 MLP rows.
|
|
730
|
+
assert bill(98) == 0.25 * GIB + 36 * mlp_row * 98
|
|
731
|
+
assert 0.45 * GIB < bill(98) < 0.65 * GIB
|
|
732
|
+
assert 0.74 * GIB < bill(195) < 0.97 * GIB
|
|
733
|
+
assert bill(4096) == 0.75 * GIB + 10 * mlp_row * 4096
|
|
732
734
|
# Fused attention: no term grows with the keys (the QSA bill would
|
|
733
735
|
# charge 11 GB of scores at 99K keys).
|
|
734
736
|
assert bill(2048, prompt_tokens=1_000) == bill(2048, prompt_tokens=262_144)
|
|
@@ -1485,8 +1487,10 @@ class Test525SixtyFourGigSeat:
|
|
|
1485
1487
|
assert growth["live_prefill_bytes"] == 250_000 * Q27_KV
|
|
1486
1488
|
assert growth["repage_copy_bytes"] == (250_000 + 16_386) * R525_Q4
|
|
1487
1489
|
assert growth["growth_bytes"] == growth["repage_bytes"]
|
|
1488
|
-
assert growth["scratch_source"] == "
|
|
1489
|
-
|
|
1490
|
+
assert growth["scratch_source"] == "geometry_measured"
|
|
1491
|
+
# The 27B's measured 2,048-row bill (2.52 to 2.72 GiB measured).
|
|
1492
|
+
assert growth["scratch_bytes"] == srv._dense_prefill_bill(_q27_text_args(), 2048)
|
|
1493
|
+
assert 2.72 * GIB < growth["scratch_bytes"] < 2.95 * GIB
|
|
1490
1494
|
# The old projection, the q4 width and the flat reserve (7.6 GiB),
|
|
1491
1495
|
# missed 12.6 GiB of it.
|
|
1492
1496
|
old_growth = 250_000 * R525_Q4 + 3 * GIB
|
|
@@ -902,6 +902,9 @@ class TestThe4B:
|
|
|
902
902
|
)
|
|
903
903
|
manager = _manager()
|
|
904
904
|
_roomy(monkeypatch, manager)
|
|
905
|
+
# The 4B's own seat: its weights and a few GiB of headroom under the
|
|
906
|
+
# 12 GiB limit, so the request is admitted at the profile's chunk.
|
|
907
|
+
_install(monkeypatch, _Machine(manager.bank, base_gib=2.6, cache_gib=0.0, host_gib=1.0))
|
|
905
908
|
state = _state(manager, plan=plan, runtime=self._runtime(), limit_gib=12, total_gib=16)
|
|
906
909
|
pricing: dict = {}
|
|
907
910
|
srv._prefill_admission_shed(
|
|
@@ -914,10 +917,17 @@ class TestThe4B:
|
|
|
914
917
|
pricing=pricing,
|
|
915
918
|
)
|
|
916
919
|
growth = pricing["growth"]
|
|
917
|
-
|
|
918
|
-
|
|
920
|
+
# A dense family: what was measured on it (2026-10-02, the 4B
|
|
921
|
+
# through mtplx serve: 1.55 to 1.65 GiB at 2,048 rows from 7K to 49K
|
|
922
|
+
# of context), 0.75 GiB plus ten MLP rows (2 x (2,560 + 3 x 9,216))
|
|
923
|
+
# a row at this width: 1.90 GiB. 2.12.1 charged the 27B's 3 GiB.
|
|
924
|
+
args = srv._runtime_text_args(self._runtime())
|
|
919
925
|
assert growth["prefill_chunk_tokens"] == 2048
|
|
920
926
|
assert growth["scratch_rows"] == 2048
|
|
921
|
-
assert growth["scratch_source"] == "
|
|
922
|
-
assert growth["scratch_bytes"] ==
|
|
927
|
+
assert growth["scratch_source"] == "geometry_measured"
|
|
928
|
+
assert growth["scratch_bytes"] == srv._dense_prefill_bill(args, 2048)
|
|
929
|
+
assert growth["scratch_bytes"] == 0.75 * GIB + 10 * 2 * (2_560 + 3 * 9_216) * 2048
|
|
930
|
+
assert 1.65 * GIB < growth["scratch_bytes"] < 1.91 * GIB
|
|
931
|
+
# A 59-token prompt (0.19 GiB measured; 2.12.1 charged 2.26 GiB).
|
|
932
|
+
assert 0.19 * GIB < srv._dense_prefill_bill(args, 59) < 0.4 * GIB
|
|
923
933
|
assert growth["publish_copy_bytes"] == 20_000 * kv
|
|
@@ -79,7 +79,9 @@ class TestTheHostAllowance:
|
|
|
79
79
|
def test_the_default_limit_keeps_the_seats_share(self):
|
|
80
80
|
assert srv._host_memory_allowance_bytes(_caps(96, "default")) == 8 * GIB
|
|
81
81
|
|
|
82
|
-
|
|
82
|
+
# Never under the 4 GiB a small seat's own model holds outside MLX
|
|
83
|
+
# (2026-10-02): a 6 GiB limit lowers the share, not the daemon's floor.
|
|
84
|
+
@pytest.mark.parametrize("limit_gib, allowance_gib", [(96, 8), (90, 7.5), (48, 4), (6, 4)])
|
|
83
85
|
def test_an_explicit_limit_bounds_it(self, limit_gib, allowance_gib):
|
|
84
86
|
assert srv._host_memory_allowance_bytes(_caps(limit_gib, "env")) == int(
|
|
85
87
|
allowance_gib * GIB
|
|
@@ -56,14 +56,19 @@ def scoring_admission(monkeypatch):
|
|
|
56
56
|
def test_default_generic_scoring_admits_the_27b_request_that_fits(scoring_admission):
|
|
57
57
|
state, client = scoring_admission
|
|
58
58
|
prompt = [7] * 8192
|
|
59
|
-
# The generation bill
|
|
59
|
+
# The generation bill at serving width reproduced the review's refusal
|
|
60
|
+
# (24.63 GiB on the 24 GiB limit at 2,048 rows, 2.12.1). With the bill
|
|
61
|
+
# measured on the 27B (2026-10-02: 1.52 GiB at 1,024 rows, 2.55 at
|
|
62
|
+
# 2,048) it runs at 1,024 rows, under the limit.
|
|
60
63
|
priced = {}
|
|
61
64
|
previous = srv._prefill_admission_shed(
|
|
62
65
|
state, prompt_ids=prompt, session_bank=None, session_id=None,
|
|
63
66
|
max_new_tokens=0, mtp_depth=0, prefill_chunk_tokens=2048, pricing=priced,
|
|
64
67
|
)
|
|
65
|
-
assert previous
|
|
66
|
-
assert previous["
|
|
68
|
+
assert not previous.get("refused")
|
|
69
|
+
assert previous["prefill_chunk_requested"] == 2048
|
|
70
|
+
assert previous["prefill_chunk_tokens"] == 1024
|
|
71
|
+
assert 23.28 * GIB < previous["projected_bytes_after"] < 24 * GIB
|
|
67
72
|
|
|
68
73
|
response = client.post("/v1/completions", json={
|
|
69
74
|
"prompt": prompt, "echo": True, "logprobs": 2, "max_tokens": 0,
|