docs: update kernelwiki using evolve skills

artifacts/blogs/flash-attention-4/MANIFEST.yaml

@@ -4,11 +4,12 @@ code_present: true
 total_blocks: 2
 generated_by: scripts/extract_blog_code.py
 files:
-- local_path: code/01-software-exp-cody-waite-horner.cu
-  heading_path: '## Key Code > ### Software exp (Cody-Waite + Horner)'
+- local_path: code/01-software-exp-published-range-reduction-and-rounded-polynomia.cu
+  heading_path: '## Illustrative Code > ### Software exp (published range reduction
+    and rounded polynomial)'
   fence_lang: cuda
-  sha256: 6fba9c537831bba7b175ee3a050e3adb5ddf99193dee550b7122a4663e3f9164
+  sha256: 89c085df18aa5ae9b8693d1085bf43881b8bc3644068761e0312c9ee7628e71c
 - local_path: code/02-2-cta-cooperative-backward.cu
-  heading_path: '## Key Code > ### 2-CTA cooperative backward'
+  heading_path: '## Illustrative Code > ### 2-CTA cooperative backward'
   fence_lang: cuda
-  sha256: e089feff40c10ebf42fa1ff878f775c21f4f044fc4b6f2f510b43c2666c7e219
+  sha256: 5747bb5ef24ec8a82d9fd80d433d0dc231d0de042af83eeec9414b03a1a53de9

artifacts/blogs/flash-attention-4/code/01-software-exp-published-range-reduction-and-rounded-polynomia.cu

@@ -0,0 +1,15 @@
+// Extracted from sources/blogs/flash-attention-4.md by scripts/extract_blog_code.py
+// Heading: ## Illustrative Code > ### Software exp (published range reduction and rounded polynomial)
+// Original fence language: cuda
+// See artifacts/blogs/flash-attention-4/code/PROVENANCE.yaml for origin + license metadata.
+
+// KernelWiki scalar illustration derived from the FA4 blog equations.
+// This is not verbatim upstream FA4 code and omits selection and clamping.
+#include <cmath>
+
+__host__ __device__ inline float fa4_blog_exp2_reference(float x) {
+    const int n = static_cast<int>(floorf(x));
+    const float f = x - static_cast<float>(n);  // f in [0, 1)
+    const float p = 1.0f + f * (0.6951f + f * (0.2276f + f * 0.0771f));
+    return ldexpf(p, n);
+}

artifacts/blogs/flash-attention-4/code/02-2-cta-cooperative-backward.cu

@@ -1,10 +1,14 @@
 // Extracted from sources/blogs/flash-attention-4.md by scripts/extract_blog_code.py
-// Heading: ## Key Code > ### 2-CTA cooperative backward
+// Heading: ## Illustrative Code > ### 2-CTA cooperative backward
 // Original fence language: cuda
 // See artifacts/blogs/flash-attention-4/code/PROVENANCE.yaml for origin + license metadata.
 
-// 2-CTA cooperative backward: paired CTAs in a cluster share a single TMEM
-// accumulator half, halving SMEM traffic for dK/dV accumulation.
-asm volatile(
-    "tcgen05.mma.cta_group::2.kind::f16 [%0], %1, %2, %3, 1;"
-    : : "r"(tmem_acc_shared), "l"(desc_a), "l"(desc_b), "r"(0));
+// KernelWiki schematic derived from the FA4 paper/blog dimensions.
+// This is not upstream inline PTX or a complete kernel.
+struct Fa4TwoCtaBackwardShape {
+    static constexpr int cta_group = 2;
+    static constexpr int mma_m = 256;
+    static constexpr int mma_n = 128;
+    static constexpr int mma_k = 128;
+    static constexpr int backward_gemm_count = 5;
+};

artifacts/blogs/flash-attention-4/code/PROVENANCE.yaml

@@ -1,21 +1,22 @@
 origin_url: https://tridao.me/blog/2026/flash4/
 upstream_repo: blog
-upstream_sha: c9a560f44002da92d82680036e3482d0ac4939a3
+upstream_sha: none
 license: inherits-from-source-blog
-retrieved_at: 2026-04-27
+retrieved_at: 2026-08-08
 asset_mode: extracted
 generated_by: scripts/extract_blog_code.py
 size_cap_truncated: false
 files:
-- local_path: 01-software-exp-cody-waite-horner.cu
+- local_path: 01-software-exp-published-range-reduction-and-rounded-polynomia.cu
   role: extracted-block
   mode: extracted
   upstream_path: inline-in-blog-markdown
-  heading_path: '## Key Code > ### Software exp (Cody-Waite + Horner)'
-  sha256: 6fba9c537831bba7b175ee3a050e3adb5ddf99193dee550b7122a4663e3f9164
+  heading_path: '## Illustrative Code > ### Software exp (published range reduction
+    and rounded polynomial)'
+  sha256: 89c085df18aa5ae9b8693d1085bf43881b8bc3644068761e0312c9ee7628e71c
 - local_path: 02-2-cta-cooperative-backward.cu
   role: extracted-block
   mode: extracted
   upstream_path: inline-in-blog-markdown
-  heading_path: '## Key Code > ### 2-CTA cooperative backward'
-  sha256: e089feff40c10ebf42fa1ff878f775c21f4f044fc4b6f2f510b43c2666c7e219
+  heading_path: '## Illustrative Code > ### 2-CTA cooperative backward'
+  sha256: 5747bb5ef24ec8a82d9fd80d433d0dc231d0de042af83eeec9414b03a1a53de9

artifacts/blogs/flashmla/MANIFEST.yaml

@@ -1,14 +1,15 @@
 slug: flashmla
-origin_url: https://github.com/deepseek-ai/FlashMLA
+origin_url: https://github.com/deepseek-ai/FlashMLA/tree/71c737929f2567bd0a094ae140f8f60f390b1232
 code_present: true
 total_blocks: 2
 generated_by: scripts/extract_blog_code.py
 files:
-- local_path: code/01-mla-decode-inner-loop.cu
-  heading_path: '## Key Code > ### MLA decode inner loop'
-  fence_lang: cuda
-  sha256: 4d3950b24b3da7a36e1db9d60a0492ed963d1b49ad6a1b97adeab54abaf4548f
-- local_path: code/02-sparse-mla-kv-retrieval-kernel-v3-2.cu
-  heading_path: '## Key Code > ### Sparse-MLA KV-retrieval kernel (V3.2)'
-  fence_lang: cuda
-  sha256: 9b0499c0ddb77a9cebd370e84929d7149ebdc03163768369c95a12ae0b5c4f7f
+- local_path: code/01-v3-fp8-sparse-decode-byte-check.py
+  heading_path: '## Exact V3 FP8 Sparse-Decode Layout > ### V3 FP8 sparse-decode byte
+    check'
+  fence_lang: python
+  sha256: bfc0a6df6b2e04ce4f894a9df04926d4120b4e370354d68e80a55c1645565c7e
+- local_path: code/02-decode-page-index-round-trip.py
+  heading_path: '## Sparse Index Contracts > ### Decode page-index round trip'
+  fence_lang: python
+  sha256: a0bbc6f0248adc67b8113e3faa66c1be8924aafce665ebf89b43f0e2f0e4a405

artifacts/blogs/flashmla/code/01-v3-fp8-sparse-decode-byte-check.py

@@ -0,0 +1,16 @@
+# Extracted from sources/blogs/flashmla.md by scripts/extract_blog_code.py
+# Heading: ## Exact V3 FP8 Sparse-Decode Layout > ### V3 FP8 sparse-decode byte check
+# Original fence language: python
+# See artifacts/blogs/flashmla/code/PROVENANCE.yaml for origin + license metadata.
+
+# KernelWiki-derived contract check; not upstream FlashMLA code.
+NOPE_FP8_VALUES = 512
+GROUPS = 4
+FP32_BYTES = 4
+ROPE_BF16_VALUES = 64
+BF16_BYTES = 2
+
+V3_FP8_SPARSE_BYTES = (
+    NOPE_FP8_VALUES + GROUPS * FP32_BYTES + ROPE_BF16_VALUES * BF16_BYTES
+)
+assert V3_FP8_SPARSE_BYTES == 656

artifacts/blogs/flashmla/code/02-decode-page-index-round-trip.py

@@ -0,0 +1,15 @@
+# Extracted from sources/blogs/flashmla.md by scripts/extract_blog_code.py
+# Heading: ## Sparse Index Contracts > ### Decode page-index round trip
+# Original fence language: python
+# See artifacts/blogs/flashmla/code/PROVENANCE.yaml for origin + license metadata.
+
+# KernelWiki-derived contract check; not upstream FlashMLA code.
+def encode_page_index(physical_page: int, offset: int, page_size: int) -> int:
+    assert physical_page >= 0 and 0 <= offset < page_size
+    return physical_page * page_size + offset
+
+def decode_page_index(encoded: int, page_size: int) -> tuple[int, int]:
+    assert encoded >= 0 and page_size > 0
+    return divmod(encoded, page_size)
+
+assert decode_page_index(encode_page_index(7, 13, 64), 64) == (7, 13)

artifacts/blogs/flashmla/code/PROVENANCE.yaml

@@ -1,21 +1,22 @@
-origin_url: https://github.com/deepseek-ai/FlashMLA
+origin_url: https://github.com/deepseek-ai/FlashMLA/tree/71c737929f2567bd0a094ae140f8f60f390b1232
 upstream_repo: blog
-upstream_sha: 71c737929f2567bd0a094ae140f8f60f390b1232
+upstream_sha: none
 license: inherits-from-source-blog
 retrieved_at: 2026-04-27
 asset_mode: extracted
 generated_by: scripts/extract_blog_code.py
 size_cap_truncated: false
 files:
-- local_path: 01-mla-decode-inner-loop.cu
+- local_path: 01-v3-fp8-sparse-decode-byte-check.py
   role: extracted-block
   mode: extracted
   upstream_path: inline-in-blog-markdown
-  heading_path: '## Key Code > ### MLA decode inner loop'
-  sha256: 4d3950b24b3da7a36e1db9d60a0492ed963d1b49ad6a1b97adeab54abaf4548f
-- local_path: 02-sparse-mla-kv-retrieval-kernel-v3-2.cu
+  heading_path: '## Exact V3 FP8 Sparse-Decode Layout > ### V3 FP8 sparse-decode byte
+    check'
+  sha256: bfc0a6df6b2e04ce4f894a9df04926d4120b4e370354d68e80a55c1645565c7e
+- local_path: 02-decode-page-index-round-trip.py
   role: extracted-block
   mode: extracted
   upstream_path: inline-in-blog-markdown
-  heading_path: '## Key Code > ### Sparse-MLA KV-retrieval kernel (V3.2)'
-  sha256: 9b0499c0ddb77a9cebd370e84929d7149ebdc03163768369c95a12ae0b5c4f7f
+  heading_path: '## Sparse Index Contracts > ### Decode page-index round trip'
+  sha256: a0bbc6f0248adc67b8113e3faa66c1be8924aafce665ebf89b43f0e2f0e4a405

artifacts/blogs/gated-delta-net/MANIFEST.yaml

@@ -1,14 +1,4 @@
 slug: gated-delta-net
-origin_url: https://github.com/NVlabs/GatedDeltaNet
-code_present: true
-total_blocks: 2
-generated_by: scripts/extract_blog_code.py
-files:
-- local_path: code/01-chunk-parallel-prefill-reference-pytorch.py
-  heading_path: '## Key Code > ### Chunk-parallel prefill reference (PyTorch)'
-  fence_lang: python
-  sha256: 709cd74d72c22e764ffb77d85bbecd0b1c47b52f2cc0c48bab279954d0f24a22
-- local_path: code/02-triton-decode-step-kernel-streaming.py
-  heading_path: '## Key Code > ### Triton decode-step kernel (streaming)'
-  fence_lang: python
-  sha256: 09a1bc93abdf8165a3d7809e485ef17a3d1a980ae579e36ba81b47665ba35b22
+origin_url: https://github.com/NVlabs/GatedDeltaNet/tree/b53d6d3a161267432a79c1c04af69fa52bddc921
+code_present: false
+generated_by: KernelWiki verifier remediation

artifacts/kernels/flash-attention-4/full/PROVENANCE.yaml

@@ -6,6 +6,10 @@ retrieved_at: '2026-04-17'
 asset_mode: verbatim
 size_cap_truncated: false
 generated_by: Phase 3 Round 2 task11 anchor bundle
+description: >-
+  Adjacent NVIDIA CUTLASS SM100 FMHA backward MLA reference. This bundle is
+  not the Dao-AILab FlashAttention-4 implementation; it is retained only as a
+  separately pinned comparison implementation.
 source_pr_id: pr-cutlass-2466
 files:
   - local_path: sm100_fmha_bwd_mla_kernel_tma_warpspecialized.hpp