diff --git a/npu/pack_import.cpp b/npu/pack_import.cpp index a306a6b88..b0f3719fd 100644 --- a/npu/pack_import.cpp +++ b/npu/pack_import.cpp @@ -393,6 +393,16 @@ constexpr Fconst19Rule kFconst19Rules[] = { // prefix as Qwen3-1.7B at this hidden size (8,512), a different bundle (5 tiles of 591,872, not 3 of 1,183,744), // and 1,857 of 686,518,336 bytes differ (2.7e-6). This row is the one that disproved the QK-norm explanation above. {2048, 32, 16, 8512, 5, 1, 591872}, + // Phi-3.5-mini-instruct at hidden 3,072 / 32 heads / 32 layers (and Phi-3-mini-128k, the same phi3 geometry; the + // table is keyed by shape, and the Llama row above is what established that the prefix is a property of the shape). + // Measured against Phi-3.5-mini's own 1.8 build (`_rai_1.8.0_npu_4K`) by whole-file normalisation: 28 of the 32 + // blocks come out byte-identical and the four that differ hold all 26,145 differing bytes of 2,036,067,648 + // (1.28e-5), i.e. content in four layers, not layout. The separation that makes this a rule rather than a fit is + // the wrong-rotation end: any rotation other than this one differs in 307,097,372 bytes (15 %) across all 32 + // blocks, a factor of 10,000. The two halves of the bundle are swapped relative to 1.8 -- the same q/k/v order + // change the transform above notes -- so n=2/rot=1 is the minimal spelling of it (n=4/rot=2 and n=8/rot=4 are the + // identical permutation, which is why the measurement tool prints the same residual for each). + {3072, 32, 32, 10560, 2, 1, 5222400}, // Mistral-7B-Instruct-v0.2 at hidden 4,096 / 32 heads / 32 layers: measured with tools/fconst19_measure.py against // amd/Mistral-7B-Instruct-v0.2_rai_1.8.0_npu_4K, 1,954 of 3,843,419,200 bytes differ (5.08e-05) -- content, not // layout. The prefix follows the Qwen3 line (64 + 2*hidden + 4,352 = 12,608), the bundle is 5 tiles of 2,297,856