bendtorch

Releases (6)

• 1.3.0.0 indexed
• 1.2.0.0 unindexed
• 1.1.1.0 unindexed
• 1.1.0.0 unindexed
• 1.0.1.0 unindexed
• 1.0.0.0 unindexed

Definitions (1,049)

main.bend

• def ts_carry
• def ts_pass
• def ts_first
• def ts_go
• def tree_sum
• type CSMCross-entropy's softmax over a row of c classes: as psm, but the row's exponentials are summed as a balanced tree (tree_sum, on a copy), since a running F32 sum of SmolLM2's 49152 classes was 3.4 tolerances off PyTorch's loss (NOTES, M10).
• def cesm4
• def cesm3
• def cesm_e
• def cesm2
• def ce_sm
• def ce_gz[idx] = (p - [j == y]) / nf for j = jj..jj + k, p the row's probabilities
• def ce_pick2loss - log_softmax(x)[y], log_softmax as PyTorch computes it: (x[y] - m) - log(sum exp(x - m)) (since 2026-10-18; it was loss - log(max(p[y], 1e-12)), which capped a row's loss at 27.63 where PyTorch's grows, NOTES)
• def ce_pick3
• def ce_bad
• def ce_picka label past the last class picks probability 0.0, clamped at 1e-12, as tensor-array's pick
• def ce_row2
• def ce_row
• def ce_rows
• def ce_fin
• def CrossEntropy
• type D3S
• def mse_lover every number from idx on (k more): d = y - t, q[idx] = d·d, y[idx] = d * kk
• def mse_fin
• def MSEmean squared error over every number (PyTorch's F.mse_loss); its gradient is 2 (y - t) / (n*c); the squares summed as a tree
• type NS
• def nll_hit
• def nll_onea label past the last class picks 0.0 and leaves its gradient row zero, as l_at and onehot do
• def nll_rowsthe sum of the picked log-probabilities (in row order, from +0.0), and the gradient (-1/n at each label)
• def nll_fin
• def NLLnegative log-likelihood of log-probabilities (PyTorch's F.nll_loss, mean over the rows)
• def byte_of
• def read_all
• def ph_hw
• def ph_bool
• type PTU
• def ptu_node
• def pt_cellsthe number of cells a tree takes (the tree comes back)
• type PTA
• type UA
• def words_outwords i..
• def words_in
• def pta_then
• def pta_node
• def pts_v
• def pts_m
• def ptsv_v
• def ptsv_m
• def pt_storethe tree's cells into xs from off
• def ptl_vreading a leaf back: fresh buffers filled from xs
• def ptl_m
• def ptlv_v
• def ptlv_m
• type PKwhat a tag says
• def pt_kind2
• def pt_kind
• def pt_nextthe tag at off and the cell after it, to a continuation
• def pt_readthe part at off, of kind k, with its second cell rn already read; fuel bounds the depth (a model's pairs nest far less than 1000 deep)
• def pt_loada tree from xs at off
• type TrainPL
• def td0
• def td1
• def td3
• def cells_lthe first k cells of an array, in order
• def tdp2
• def tdp
• def Train.dumpevery number of a training state: the parameters, then the first moments, the second moments, b1^t and b2^t (packed moments: their tree's cells, pt_store's layout, in place of the two moments)
• def f32_at
• def tr0
• def tr_m2
• def tr_m1
• def tr_adam2
• def tr_adam
• def tr3
• def tr2
• def tr1
• def trp4
• def trp3
• def trp2
• def trp
• def Train.restorea training state filled from the numbers Train.dump wrote (its buffers are overwritten in place; packed moments are built again from the tree's cells)
• def WCHUNK
• def chunk_bytesthe bytes of xs[i - k + 1 ..
• def wc_ok
• def wc_bytes
• def write_chunksk chunks of a's numbers lo ..
• def wf_close
• def write_f32cells 0 ..
• def ps_fin
• def Params.saveevery parameter of p, in dump order, into a file at path (Layer.count numbers), through one array
• def tsa_fin
• def tsa_m
• def tsa_pw
• def tsa_v
• def tsa_mm
• def tsa_mv
• def tsp3
• def tsp2
• def tsp
• def Train.savethe layout Train.dump gives, written from one array instead of a list (the same bytes)