3 people 171acceded [Bugfix][SM70] GDN prefill: use the prebuilt kernel when TileLang has no nvcc (#1400) ...

* [Bugfix][SM70] GDN prefill: use the prebuilt kernel when TileLang has no nvcc

The original FlashQLA prefill is the default on SM70 and JIT-compiles
with TileLang on first use. A bare wheel install has no CUDA toolkit, so
the first prefill failed with "No CUDA or HIP or MPS available on this
system" (or later, when /usr/local/cuda holds only the runtime).

When neither VLLM_SM70_FLASHQLA_ORIGINAL_PREFILL nor its legacy alias is
set and TileLang would find no nvcc, the default now falls back to the
prebuilt VLK prefill and logs once. An explicit request still selects the
original path.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* [Doc] Regenerate the runtime architecture reference for gdn.py

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: mac <mac@MacBook-xinyi.local>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

2026-11-06 21:57:50 +08:00

..

architecture

[Bugfix][SM70] GDN prefill: use the prebuilt kernel when TileLang has no nvcc (#1400)

2026-11-06 21:57:50 +08:00

data

[Core] Integrate main policies and isolate HC feature callbacks

2026-10-09 18:23:23 +08:00

minimax_h3

…

native_media

…

arch_overview.md

…

attention_backends.md

…

cuda_graphs_multimodal.md

…

cuda_graphs.md

…

custom_op.md

…

dbo.md

…

debug_vllm_compile.md

…

early_graph_model_inputs.md

…

environment_metadata.md

…

flashnext_16gb_host_kv.md

[Core] Share capture streams across serial speculative graph routines

2026-10-11 10:37:58 +08:00

flashnext_abliterated_9683_20261014.md

[Deploy] Qualify the abliterated model on 9683

2026-10-14 22:20:17 +08:00

flashnext_abliterated_local_20261014.md

[Deploy] Qualify the abliterated model on 9683

2026-10-14 22:20:17 +08:00

flashnext_decode_verifier_16gb_20261014.md

[Perf] Keep encoded target KV in the Flash-Next hot cache

2026-10-14 12:26:25 +08:00

flashnext_device_qsa_history.md

[Doc] Record installed-wheel direct QSA validation

2026-10-09 20:07:07 +08:00

flashnext_gguf_graph_parity.md

…

flashnext_hcx_local_schedule.md

[Test] Validate native HCXO scheduling and record structural screens

2026-10-10 03:48:48 +08:00

flashnext_host_fp8_kv.md

[Doc] Record matched E4M3 KV placement measurements

2026-10-08 00:19:44 +08:00

flashnext_indexer_placement_20261015.md

[Kernel] Reuse exact QSA selectors over decode score partitions

2026-10-15 05:25:37 +08:00

flashnext_mtp4_latency_20261008.md

[Doc] Diagnose current Flash-Next MTP4 round latency

2026-10-08 02:29:27 +08:00

flashnext_mtp4_round12_screens_20261008.md

[KV] Integrate main policies and resolve history behind configuration

2026-10-09 18:25:52 +08:00

flashnext_mtp4_round12.md

[KV] Skip initialization of fully written host QSA rows

2026-10-08 04:40:00 +08:00

flashnext_mtp4_structural_costs.md

[Profile] Add an opt-in SM70 round-cost ledger (#1254)

2026-10-24 00:45:45 +08:00

flashnext_persistent_execution.md

[Profile] Add an opt-in SM70 round-cost ledger (#1254)

2026-10-24 00:45:45 +08:00

flashnext_qsa_router_structure.md

[Kernel] Reduce exact QSA selection work for long contexts (#1171)

2026-10-11 04:27:05 +08:00

fused_moe_modular_kernel.md

…

fusions.md

…

gguf_collective_norm.md

[Doc] Record matched GGUF collective norm model results

2026-10-07 02:02:32 +08:00

gguf_dense_segments.md

…

gguf_dense_weight_reuse.md

[Doc] Record clean native validation and rejected deeper prefetch

2026-10-10 02:57:23 +08:00

gguf_dflash2_iq3s.md

…

gguf_dflash2_projection_ledger.md

…

gguf_dflash2_tp_sizes.md

[Core] Share DFlash attention and norm boundaries across TP sizes (#1025)

2026-10-07 15:10:21 +08:00

gguf_dmv_projections.md

[Doc] Separate projection increments from graph-context costs

2026-10-06 23:05:38 +08:00

gguf_dmv_tp2.md

[Kernel] Share GGUF projection planes with TP2 shards (#1026)

2026-10-07 16:38:06 +08:00

gguf_dp4a_small_m.md

…

gguf_fp16_projection.md

…

gguf_fp16_short_context_splits.md

[Kernel] Improve FP16 grouped verification at short contexts (#1022)

2026-10-07 12:07:00 +08:00

gguf_head_layout_preparation.md

…

gguf_iq2_signed_nibbles.md

[Doc] Qualify IQ2 projection gains in complete rounds

2026-10-07 07:18:55 +08:00

gguf_iq3_gated_pair_dispatch.md

…

gguf_iq3_xxs_pair_reader.md

…

gguf_iq3_xxs_pair_records.md

…

gguf_iq3xxs_joint_experts.md

…

gguf_iq4_native_reader.md

…

gguf_lm_head_sm70.md

…

gguf_lut4_q8_experts.md

[Doc] Record matched IQ4 model quality and partition screening

2026-10-06 05:47:14 +08:00

gguf_mixed_projection_output.md

…

gguf_native_iq1_m_pair.md

…

gguf_native_iq2_s_pair.md

…

gguf_native_iq2_xs_pair.md

…

gguf_native_iq2_xxs_pair.md

…

gguf_native_iq3xxs_iq4_pair.md

…

gguf_native_mixed_pair.md

…

gguf_native_q2_k_pair.md

…

gguf_native_q4_iq3_pair.md

…

gguf_native_q4_k_pair.md

…

gguf_native_sm70.md

…

gguf_ordered_ple_rows.md

…

gguf_packed_embeddings.md

[Doc] Record packed GGUF embedding correctness and storage

2026-10-08 00:33:39 +08:00

gguf_packed_ple_results.md

[Core] Retain owned FP16 PLE result storage across replay

2026-10-09 07:57:57 +08:00

gguf_packed_ple.md

…

gguf_pinned_ple_decode.md

…

gguf_projection_clock_diagnostics.md

[Kernel] Add device warp timing for GGUF projection diagnostics (#1018)

2026-10-07 10:49:56 +08:00

gguf_q4_k_pair_records.md

…

gguf_q8_intermediate.md

…

gguf_qkv_profile_calibration.md

[Kernel] Fuse measured three-format GGUF QKV projections (#1017)

2026-10-07 09:24:37 +08:00

gguf_qpn_attention_inputs.md

[Doc] Record installed GGUF attention projection gates

2026-10-06 04:24:15 +08:00

gguf_qpn_canonical_streams.md

[Doc] Retain canonical QPN screening and reject slower routes

2026-10-06 04:34:03 +08:00

gguf_qpn_linear.md

…

gguf_qpn_output_batch.md

[Doc] Add same-clock head reference and reporting checks

2026-10-06 05:24:52 +08:00

gguf_qpn_projection_batch.md

…

gguf_qpn_qkvz.md

…

gguf_qpn_register_pair.md

[Doc] Preserve the rejected IQ3 register-pair screen

2026-10-09 09:47:08 +08:00

gguf_qpn_small_outputs.md

…

gguf_qwen35_moe_adapter.md

…

gguf_qwen38_iq3s_layer_types.csv

…

gguf_qwen38_iq3s_source_inventory.md

…

gguf_qwen38_mixed_pair_coverage.md

…

gguf_tp2_256k_memory.md

[Doc] Record dual V100 GGUF DFlash2 256K execution

2026-10-08 00:35:04 +08:00

gguf_tp4_verify_cost.md

[Doc] Measure TP4 GGUF DFlash2 verification cost

2026-10-08 03:07:13 +08:00

gguf_turbomind_bitplanes.md

…

gguf_turbomind_codebook_words.md

…

gguf_turbomind_iq3_packet.md

…

gguf_turbomind_lattice_banks.md

…

gguf_turbomind_lattice_decode.md

…

gguf_turbomind_lattice_grouped.md

…

gguf_turbomind_lattice.md

…

gguf_turbomind_lut4.md

…

gguf_turbomind_model_moe.md

…

gguf_turbomind_model.md

…

gguf_turbomind_prefill.md

[Perf] Bring Codex's 9683 prefill campaign work onto the release line

2026-10-20 06:03:24 -07:00

gguf_turbomind_u2.md

…

gguf_turbomind.md

…

gguf_v100_control.md

[Model] Integrate standalone Qwen4Exp GGUF mapping on main

2026-10-05 00:47:24 +08:00

grouped_cpu_offload.md

[Bugfix] Reject mismatched filesystem KV page sizes (#1170)

2026-10-11 03:35:45 +08:00

h3_encoder_prepared_cache.md

…

huggingface_integration.md

…

hybrid_kv_cache_manager.md

…

io_processor_plugins.md

…

logits_processors.md

…

lora_resolver_plugins.md

…

metrics.md

…

mixed_prefill_latency.md

…

mm_processing.md

…

model_runner_v2.md

…

moe_kernel_features.md

…

multiprocessing.md

…

nixl_kv_cache_lease.md

…

omni_workflow_coverage.md

…

optimization_levels.md

…

p2p_nccl_connector.md

…

paged_attention.md

…

ple_mapped_row_gather.md

…

ple_metadata_request_indices.md

…

ple_split_mainline_audit.md

…

plugin_system.md

…

prefix_caching.md

…

qwen4exp_ple_tier_cascade.md

…

qwen4exp_qsa_dcp2.md

…

qwen36_27b_awq_mtp_target_verifier_fastpaths.md

…

safetensors_direct_io.md

…

sm70_awq_exact_m5_batched_gemv.md

…

sm70_awq_fp8_mtp_experts.md

…

sm70_awq_long_prefill_exact_dense.md

…

sm70_awq_qpn_m1.md

…

sm70_awq_small_n_hmma_operator.md

…

sm70_compact_candidate_gather.md

[Doc] Record compact candidate gather measurements

2026-10-08 06:56:23 +08:00

sm70_concurrency_integration_audit.md

…

sm70_deepseek_v4_dspark_acceptance.md

…

sm70_deepseek_v4_dspark_n7_latest_main.md

…

sm70_deepseek_v4_dspark_optimization.md

…

sm70_deepseek_v4_dspark_verifier_20ms_long_context.md

…

sm70_deepseek_v4_flash_bringup.md

…

sm70_deepseek_v4_fp16_gemv.md

…

sm70_deepseek_v4_mhc_decode.md

…

sm70_deepseek_v4_mxfp4_active_expert.md

…

sm70_deepseek_v4_mxfp4_grouped_decode.md

…

sm70_deepseek_v4_performance_integration.md

…

sm70_deepseek_v4_sparse_mla_qk_dsplit.md

…

sm70_deepseek_v4_sparse_mla_splitk.md

…

sm70_deepseek_v4_tp8_graph_allreduce.md

…

sm70_dflash2_acceptance_20260909.md

…

sm70_dflash2_acceptance_rd.md

…

sm70_dflash2_attention_resources_20260910.md

…

sm70_dflash2_batch_attention_gdn.md

…

sm70_dflash2_batch_latency_20260926.md

…

sm70_dflash2_context_cost_20260909.md

…

sm70_dflash2_fastpath_numerics.md

…

sm70_dflash2_fp8_verify_20ms.md

…

sm70_dflash2_fp32_defaults.md

[Bugfix] Unify DFlash2 E4M3 attention on FP32 state

2026-09-08 09:02:17 +08:00

sm70_dflash2_long_batch_20260927.md

[Doc] Record DFlash2 merge disposition and remaining gates

2026-09-27 02:11:13 +08:00

sm70_dflash2_long_decode_followup_20260927.md

…

sm70_dflash2_long_verify_curve_20260909.md

…

sm70_dflash2_long_verify_decay.md

…

sm70_dflash2_lookup_augmented.md

…

sm70_dflash2_m48_scaling_20260927.md

…

sm70_dflash2_ngram_hybrid.md

…

sm70_dflash2_nvfp4_17ms.md

…

sm70_dflash2_nvfp4_prefill_promotion.md

…

sm70_dflash2_policy.md

…

sm70_dflash2_prefill_closure.md

…

sm70_dflash2_quality_audit.md

…

sm70_dflash2_retired_candidate_order.md

…

sm70_dflash2_shared_nvfp4.md

…

sm70_dflash2_tail_graphs_20260911.md

…

sm70_dflash2_target_graph_20ms.md

…

sm70_dflash2_tp2_verifier.md

…

sm70_dflash2_verifier_route_audit.md

…

sm70_dflash_ddtree_27b_awq_plan.md

…

sm70_dflash_page832.md

[Doc] Record qualified DFlash window split speed and quality

2026-10-07 04:43:45 +08:00

sm70_dflash_page1648.md

[Core] Keep DFlash split windows with E4M3 hybrid pages (#1024)

2026-10-07 12:53:06 +08:00

sm70_dflash_smallm.md

…

sm70_e4m3_grouped_fp32.md

…

sm70_environment_surface.md

…

sm70_exact_auxiliary_fusion.md

…

sm70_fa2_d256_prefill_pipeline.md

[Perf][SM70] Split-KV for short prefix-prefill chunks over long prefixes (#1298)

2026-10-29 01:56:13 +08:00

sm70_flash_next_batch_defaults.md

…

sm70_flash_next_mtp4_15ms.md

…

sm70_flash_next_mtp4_256k_eval_20260927.csv

…

sm70_flash_next_mtp4_256k_eval_20260927.md

…

sm70_flash_next_mtp4_batch_gdn.md

…

sm70_flash_next_mtp4_default_profile.md

[Release] Bring the bundled V100 launchers up to the 1.5.2 profiles (#1397)

2026-11-06 16:56:50 +08:00

sm70_flash_next_mtp4_trace_reconciliation.md

…

sm70_flash_next_prefix_prefill.md

…

sm70_flash_v37_prefill.md

…

sm70_flash_v100_fp8_kv_long_context.md

…

sm70_flash_v100_prefill_operator_optimization.md

…

sm70_flashinfer_native_backend.md

…

sm70_flashnext_gguf_mtp4_operator_routes.md

[Benchmark] Record reproducible Flash-Next GGUF MTP4 composition

2026-10-05 06:13:17 +08:00

sm70_flashnext_gguf_mtp4.md

…

sm70_flashnext_hc_boundary.md

[Kernel] Pipeline TP4 HC loads with exact FP32 reductions

2026-10-06 23:56:41 +08:00

sm70_fp8_kernel_integration.md

…

sm70_fp8_long_prefill_exact_dense.md

…

sm70_fp8_policy_cleanup.md

…

sm70_fp16_gemv_silu_schedule.md

…

sm70_fp16_grouped_verify.md

…

sm70_gdn_ba_verify.md

…

sm70_gdn_prefill_threads.md

[Perf] Bring Codex's 9683 prefill campaign work onto the release line

2026-10-20 06:03:24 -07:00

sm70_gdn_preprocess.md

…

sm70_gdn_projection_copy_audit.md

…

sm70_gdn_projection_tails.md

…

sm70_gdn_verify_tile.md

…

sm70_gemma_long_prefill_fusion.md

…

sm70_gguf_canonical_head.md

…

sm70_gguf_expert_down_dispatch.md

…

sm70_gguf_fp16_dense_routes.md

…

sm70_gguf_joint_expert_dispatch.md

…

sm70_gguf_joint_original_experts.md

…

sm70_gguf_prefill_preparation.md

[Perf] Bring Codex's 9683 prefill campaign work onto the release line

2026-10-20 06:03:24 -07:00

sm70_gguf_small_grouped_down.md

…

sm70_glm53_flash_nvfp4.md

…

sm70_hc_batch_topology.md

…

sm70_hc_ll_shards.md

…

sm70_hc_weight_prefetch_screen.md

…

sm70_hmma_pipeline_search.md

…

sm70_linear_kernel_integration.md

…

sm70_loaded_route_reporting.md

…

sm70_m8_exact_decode.md

[Kernel] Reuse M8 gate operands and TP4 norm packets (#1041)

2026-10-08 08:50:20 +08:00

sm70_mamba_state_grid_decoupling.md

…

sm70_memory_arena.md

…

sm70_memory_defaults.md

…

sm70_memory_workspace_tuning.md

…

sm70_mhc_prenorm_staging.md

[Kernel] Reduce SM70 mHC projection and output overhead (#1172)

2026-10-11 12:40:29 +08:00

sm70_mixed_batch_decode_rows.md

…

sm70_mrv2_sampling_contract_repairs.md

…

sm70_mtp_gdn_output.md

…

sm70_mtp_strided_qkv.md

…

sm70_nvfp4_bundled_scale.md

[Kernel] Record bundled NVFP4 serving validation

2026-10-07 00:49:24 +08:00

sm70_nvfp4_turbomind_operator_optimization.md

…

sm70_ple_input_prepare.md

…

sm70_ple_ngram_batch.md

…

sm70_prefill_score_overflow.md

…

sm70_prefill_tail_intermediate.md

…

sm70_prefix_anchored_swa.md

…

sm70_prefix_peak_repair.md

…

sm70_qsa_page4_allocation_invariance.md

…

sm70_quantized_batch_reuse.md

…

sm70_quasar_dflash2_15ms.md

…

sm70_quasar_dflash2_operator_audit.md

…

sm70_quasar_dflash2_quality_speed_recovery.md

…

sm70_quasar_dflash2_resource_audit_20260909.md

…

sm70_quasar_e4m3_fp32_logits.md

…

sm70_quasar_nvfp4_dflash2_acceptance.md

…

sm70_quasar_tp2_tp4_quality.md

…

sm70_qwen36_27b_awq_mtp4_optimization.md

…

sm70_qwen38_130_acceptance.md

…

sm70_qwen38_awq_active_grouped_decode.md

…

sm70_qwen38_concurrency40.md

…

sm70_qwen38_default_fastpath.md

…

sm70_qwen38_distribution_acceptance.md

…

sm70_qwen38_exact_decode_followup.md

…

sm70_qwen38_flash_next_nvfp4.md

…

sm70_qwen38_fp8_prefill_5500.md

…

sm70_qwen38_fp8_prefill_decay.md

…

sm70_qwen38_nomtp_concurrency.md

…

sm70_qwen38_nvfp4_decode.md

…

sm70_qwen38_nvfp4_quality_repair.md

…

sm70_qwen38_ple_disk_offload.md

…

sm70_qwen38_projection_w13_followup.md

…

sm70_qwen38_qpn8_decode.md

…

sm70_qwen38_reproducible_baseline.md

…

sm70_qwen38_step0_kernel_budget.md

…

sm70_release151_compile_cache.md

…

sm70_router_operators.md

[Kernel] Fuse SM70 router top-k with Q8 activation quantization (#1147)

2026-10-10 03:57:28 +08:00

sm70_sampling_input_guards.md

…

sm70_sm120_long_context_attention_parity.md

…

sm70_small_moe_routing.md

…

sm70_sparse_bmm_capabilities.md

…

sm70_tile_runtime_exploration.md

…

sm70_tp4_mtp4_multigpu_scaling.md

…

sm70_tp4_nomtp_long_context_decode.md

…

sm70_tp4_push_gemma_norm.md

…

sm70_tp_quality_audit.md

[Test] Record matched quality audits and cutoff coverage

2026-09-22 00:44:17 +08:00

sm70_tp_shape_coverage.md

[Doc] Record paired quality audit and QPN2 precision repair

2026-09-21 22:30:09 +08:00

sm70_v100_migration_control.md

[Chore] Remove research-only benchmark artifacts

2026-10-11 03:05:42 +08:00

speculative_draft_load_config.md

…

torch_compile_multimodal.md

…

torch_compile.md

…

vllm_ir.md

…