wip: T10 batched forward (validation)

crates/xtrain-autodiff/src/ops.rs

@@ -120,15 +120,17 @@ pub fn swiglu(gate: &Var, up: &Var) -> Var {
     mul(&silu(gate), up)
 }
 
-/// RoPE (rotate_half) over `x:[tokens,heads,head_dim]`. Orthogonal map, so the
+/// RoPE (rotate_half) over `x:[tokens,heads,head_dim]` with per-sequence position
-/// backward is the inverse rotation of `dy` — no cached forward values needed.
+/// `row % period` (`period` = sequence length; `period == tokens` for a single
-pub fn rope(x: &Var, theta: f32) -> Var {
+/// sequence). Orthogonal map, so the backward is the inverse rotation of `dy` — no
-    let out = x.value().rope(theta);
+/// cached forward values needed.
+pub fn rope(x: &Var, theta: f32, period: usize) -> Var {
+    let out = x.value().rope(theta, period);
     Var::from_op(
         out,
         vec![x.clone()],
         Box::new(move |dy, parents| {
-            Var::push_grad(&parents[0], Tensor::rope_backward(dy, theta));
+            Var::push_grad(&parents[0], Tensor::rope_backward(dy, theta, period));
         }),
     )
 }

crates/xtrain-autodiff/tests/autograd.rs

@@ -327,12 +327,12 @@ fn rope_bwd() {
     let w = fill(n, 82);
 
     let x = Var::leaf(cuda(&x_h, &[tokens, heads, head_dim]));
-    let out = ops::rope(&x, theta);
+    let out = ops::rope(&x, theta, tokens);
     scalar_loss(&out, &w).backward();
 
     let dx = x.grad().unwrap().to_device(Device::Cpu);
     let wf = w.clone();
-    let lx = move |v: &[f32], s: &[usize]| weighted_sum(&cuda(v, s).rope(theta), &wf);
+    let lx = move |v: &[f32], s: &[usize]| weighted_sum(&cuda(v, s).rope(theta, tokens), &wf);
     report(
         "rope dX",
         &grad_check(
@@ -345,6 +345,38 @@ fn rope_bwd() {
     );
 }
 
+// ---- rope batched (per-sequence position = row % period) ----
+// tokens = B*S laid end to end; period = S. Sequences 2 and 3 re-use positions
+// 0..S, so the kernel's `tok % period` must reset RoPE per sequence.
+#[test]
+fn rope_batched_bwd() {
+    require_gpu();
+    let (b, s, heads, head_dim) = (3, 4, 2, 8);
+    let tokens = b * s;
+    let n = tokens * heads * head_dim;
+    let theta = 10000.0;
+    let x_h = fill(n, 83);
+    let w = fill(n, 84);
+
+    let x = Var::leaf(cuda(&x_h, &[tokens, heads, head_dim]));
+    let out = ops::rope(&x, theta, s);
+    scalar_loss(&out, &w).backward();
+
+    let dx = x.grad().unwrap().to_device(Device::Cpu);
+    let wf = w.clone();
+    let lx = move |v: &[f32], sh: &[usize]| weighted_sum(&cuda(v, sh).rope(theta, s), &wf);
+    report(
+        "rope batched dX",
+        &grad_check(
+            &x_h,
+            &[tokens, heads, head_dim],
+            &lx,
+            dx.as_slice::<f32>(),
+            cfg_linear(),
+        ),
+    );
+}
+
 // ---- softmax ----
 #[test]
 fn softmax_bwd() {

crates/xtrain-cuda/src/ffi.rs

@@ -125,7 +125,9 @@ unsafe extern "C" {
     pub fn launch_silu_f32(x: *const f32, y: *mut f32, n: i32, s: CudaStream);
     pub fn launch_silu_dx_f32(x: *const f32, dy: *const f32, dx: *mut f32, n: i32, s: CudaStream);
 
-    // RoPE (rotate_half), x:[tokens,heads,head_dim], position = token index.
+    // RoPE (rotate_half), x:[tokens,heads,head_dim], position = (token index %
+    // period). `period` = sequence length, so a flattened batch of sequences gets
+    // per-sequence positions; period == tokens reproduces the single-sequence case.
     pub fn launch_rope_f32(
         x: *const f32,
         y: *mut f32,
@@ -133,6 +135,7 @@ unsafe extern "C" {
         heads: i32,
         head_dim: i32,
         theta: f32,
+        period: i32,
         s: CudaStream,
     );
     pub fn launch_rope_dx_f32(
@@ -142,6 +145,7 @@ unsafe extern "C" {
         heads: i32,
         head_dim: i32,
         theta: f32,
+        period: i32,
         s: CudaStream,
     );
 

crates/xtrain-distributed/src/ddp.rs

@@ -17,7 +17,7 @@ use std::thread;
 use std::time::Instant;
 
 use xtrain_autodiff::tape::Var;
-use xtrain_model::{Config, TinyTransformer, ids_tensor};
+use xtrain_model::{Config, TinyTransformer, batched_ids_tensor};
 use xtrain_optim::GpuAdamW;
 use xtrain_tensor::Device;
 use xtrain_train::checkpoint;
@@ -91,10 +91,11 @@ pub fn train_rank(
     let mut losses = Vec::with_capacity(cfg.steps);
     let mut evals = Vec::new();
     let mut best_val: Option<f32> = None;
-    // Each rank reaches the global batch mean as (Σ_global / world) · (1/b_local),
+    // Each rank runs ONE batched forward over its b_local = batch_size/world
-    // where b_local = batch_size / world (see DdpContext::all_reduce_average_grads).
+    // sequences → backward grad = local mean (Σ_local / b_local). all_reduce_average
+    // (sum across ranks, /world) then gives Σ_global/(world·b_local) = Σ_global/
+    // B_global — already the global-batch mean — so the clip pre-scale is 1.0.
     let batch_local = cfg.batch_size / ctx.world;
-    let inv_batch_local = 1.0 / batch_local as f32;
     let start = Instant::now();
     let mut tokens_seen: u64 = 0;
     // Rank 0 owns the held-out eval + best-val checkpoint (params are identical
@@ -105,31 +106,36 @@ pub fn train_rank(
         let lr = cfg.schedule.lr(step);
 
         // Draw the whole global batch from the shared RNG (same on every rank);
-        // run forward+backward only on this rank's shard. The tape SUMs the
+        // collect only this rank's shard (global index % world == rank) and run it
-        // shard's grads; the union of shards == the single-GPU batch.
+        // as ONE batched forward/backward. The union of shards == the single-GPU
-        let mut local_loss_sum = 0.0f32;
+        // batch; each rank's backward yields its local mean (Σ_local / b_local).
+        let mut inputs = Vec::with_capacity(batch_local);
+        let mut targets_v = Vec::with_capacity(batch_local);
         for i in 0..cfg.batch_size {
             let (input, target) = corpus.sample(cfg.seq_len, &mut rng);
-            if i % ctx.world != ctx.rank {
+            if i % ctx.world == ctx.rank {
-                continue; // not this rank's sequence
+                inputs.push(input);
+                targets_v.push(target);
             }
-            let ids = ids_tensor(&input, device);
+        }
-            let targets = ids_tensor(&target, device);
+        let ids = batched_ids_tensor(&inputs, device);
-            let loss = model.loss(&ids, &targets);
+        let targets = batched_ids_tensor(&targets_v, device);
-            local_loss_sum += read_scalar(&loss);
+        let loss = model.loss_batched(&ids, &targets, batch_local);
+        let local_mean = read_scalar(&loss); // Σ_local / b_local
         loss.backward();
-            tokens_seen += cfg.seq_len as u64;
+        tokens_seen += (batch_local * cfg.seq_len) as u64;
-        }
 
-        // AllReduce(sum) + /world the grads → every rank holds Σ_global/world.
+        // AllReduce(sum) + /world the grads → every rank holds Σ_global/B_global
+        // (local means summed over ranks, /world = global mean). See note above.
         ctx.all_reduce_average_grads(&params);
-        // The reported loss is the global mean: average local sums across ranks.
+        // Reported loss = global mean: sum the per-rank local sums (= mean·b_local)
-        let step_loss = all_reduce_loss(ctx, local_loss_sum) / cfg.batch_size as f32;
+        // across ranks, /B_global. With equal b_local this is mean over ranks.
+        let step_loss =
+            all_reduce_loss(ctx, local_mean * batch_local as f32) / cfg.batch_size as f32;
         losses.push(step_loss);
 
-        // clip pre_scale = 1/b_local finishes the average to Σ_global/B_global,
+        // Grads are already the global-batch mean — just clip (pre-scale 1.0).
-        // identical to the single-GPU clip(pre_scale = 1/B_global).
+        let gnorm = clip_grad_norm_gpu(&params, cfg.max_grad_norm, 1.0);
-        let gnorm = clip_grad_norm_gpu(&params, cfg.max_grad_norm, inv_batch_local);
         opt.step(lr, &params);
         for p in &params {
             p.zero_grad();

crates/xtrain-distributed/tests/ddp_correctness.rs

@@ -13,7 +13,7 @@ use std::time::Instant;
 
 use xtrain_cuda::device;
 use xtrain_distributed::{DdpConfig, DdpContext, build_model, get_unique_id, launch, train_rank};
-use xtrain_model::{Config, ids_tensor};
+use xtrain_model::{Config, batched_ids_tensor};
 use xtrain_optim::GpuAdamW;
 use xtrain_tensor::Device;
 use xtrain_train::clip::clip_grad_norm_gpu;
@@ -47,22 +47,25 @@ fn run_single_gpu(cfg: Config, corpus: &Corpus, dcfg: &DdpConfig) -> (Vec<f32>,
     let params = model.params();
     let mut opt = GpuAdamW::new(dcfg.weight_decay);
     let mut rng = dcfg.seed;
-    let inv_batch = 1.0 / dcfg.batch_size as f32;
     let mut losses = Vec::new();
 
     for step in 0..dcfg.steps {
         let lr = dcfg.schedule.lr(step);
-        let mut loss_sum = 0.0f32;
+        // Sample the whole global batch and run it as ONE batched forward/backward
+        // (matches the T10 DDP path: backward yields the global-batch mean grad).
+        let mut inputs = Vec::with_capacity(dcfg.batch_size);
+        let mut targets_v = Vec::with_capacity(dcfg.batch_size);
         for _ in 0..dcfg.batch_size {
             let (input, target) = corpus.sample(dcfg.seq_len, &mut rng);
-            let ids = ids_tensor(&input, device);
+            inputs.push(input);
-            let targets = ids_tensor(&target, device);
+            targets_v.push(target);
-            let loss = model.loss(&ids, &targets);
-            loss_sum += loss.value().to_device(Device::Cpu).as_slice::<f32>()[0];
-            loss.backward();
         }
-        losses.push(loss_sum * inv_batch);
+        let ids = batched_ids_tensor(&inputs, device);
-        clip_grad_norm_gpu(&params, dcfg.max_grad_norm, inv_batch);
+        let targets = batched_ids_tensor(&targets_v, device);
+        let loss = model.loss_batched(&ids, &targets, dcfg.batch_size);
+        losses.push(loss.value().to_device(Device::Cpu).as_slice::<f32>()[0]);
+        loss.backward();
+        clip_grad_norm_gpu(&params, dcfg.max_grad_norm, 1.0);
         opt.step(lr, &params);
         for p in &params {
             p.zero_grad();

crates/xtrain-model/src/lib.rs

@@ -24,4 +24,4 @@ pub use config::Config;
 #[cfg(not(no_cuda))]
 mod model;
 #[cfg(not(no_cuda))]
-pub use model::{TinyTransformer, ids_tensor, param_to_host};
+pub use model::{TinyTransformer, batched_ids_tensor, ids_tensor, param_to_host};

crates/xtrain-model/src/model.rs

@@ -106,16 +106,35 @@ impl TinyTransformer {
     }
 
     /// Forward over a single sequence of token `ids` (`[seq]` I32 on this
-    /// model's device). Returns the logits [`Var`] of shape `[seq, vocab]`.
+    /// model's device). Returns the logits [`Var`] of shape `[seq, vocab]`. This
+    /// is the batch-1 special case of [`forward_batched`](Self::forward_batched)
+    /// (used by the autoregressive sampler / inference path).
     pub fn forward(&self, ids: &Tensor) -> Var {
-        let seq = ids.shape()[0];
+        self.forward_batched(ids, 1)
+    }
+
+    /// Batched forward over `batch` sequences of equal length `seq`, flattened to
+    /// `[batch*seq]` I32 ids in sequence-major order (sequence 0's `seq` tokens,
+    /// then sequence 1's, …). Returns logits `[batch*seq, vocab]` in the SAME flat
+    /// layout. The whole graph runs on the flattened tokens so every linear
+    /// projection is ONE big `[batch*seq, dim] × [dim, out]` GEMM (the
+    /// GPU-filling win); only attention is sequence-aware (per-sequence causal
+    /// mask + RoPE position, NO cross-sequence attention).
+    pub fn forward_batched(&self, ids: &Tensor, batch: usize) -> Var {
+        let total = ids.shape()[0];
+        assert_eq!(
+            total % batch,
+            0,
+            "ids len {total} not divisible by batch {batch}"
+        );
+        let seq = total / batch;
         let mask = self.causal_mask(seq);
 
-        let mut h = ops::embedding(&self.embed, ids); // [seq, dim]
+        let mut h = ops::embedding(&self.embed, ids); // [batch*seq, dim]
         for b in &self.blocks {
             // --- Attention sub-block (pre-norm + residual) ---
             let normed = ops::rms_norm(&h, &b.attn_norm, self.cfg.eps);
-            let attn = self.attention(b, &normed, &mask, seq);
+            let attn = self.attention(b, &normed, &mask, batch, seq);
             h = ops::add(&h, &attn);
 
             // --- MLP sub-block (pre-norm + residual) ---
@@ -125,7 +144,7 @@ impl TinyTransformer {
         }
 
         let h = ops::rms_norm(&h, &self.final_norm, self.cfg.eps);
-        ops::matmul(&h, &self.lm_head) // [seq, vocab]
+        ops::matmul(&h, &self.lm_head) // [batch*seq, vocab]
     }
 
     /// Cross-entropy mean loss of `forward(ids)` against `targets` (`[seq]` I32).
@@ -134,57 +153,93 @@ impl TinyTransformer {
         ops::cross_entropy(&logits, targets)
     }
 
-    /// Multi-head causal self-attention. `x`:[seq,dim] (already normed).
+    /// Batched cross-entropy mean loss: `forward_batched(ids, batch)` against
-    fn attention(&self, b: &Block, x: &Var, mask: &Var, seq: usize) -> Var {
+    /// flat `targets` (`[batch*seq]` I32, same sequence-major layout). The CE mean
+    /// is over all `batch*seq` rows — identical to averaging the per-sequence
+    /// losses, so the loss value matches the looped single-sequence path.
+    pub fn loss_batched(&self, ids: &Tensor, targets: &Tensor, batch: usize) -> Var {
+        let logits = self.forward_batched(ids, batch);
+        ops::cross_entropy(&logits, targets)
+    }
+
+    /// Multi-head causal self-attention over a flattened batch. `x`:[batch*seq,dim]
+    /// (already normed), laid out sequence-major. The Q/K/V/O projections are big
+    /// `[batch*seq, dim]` GEMMs; only the scaled-dot-product attention is
+    /// sequence-aware — each (head, sequence) attends within its own `[seq,seq]`
+    /// causal window, with NO cross-sequence attention, and RoPE positions reset
+    /// per sequence (`period = seq`).
+    fn attention(&self, b: &Block, x: &Var, mask: &Var, batch: usize, seq: usize) -> Var {
         let (nh, hd) = (self.cfg.n_heads, self.cfg.head_dim);
+        let total = batch * seq;
         let scale = 1.0 / (hd as f32).sqrt();
 
-        // Project, then lay out as per-head [seq, head_dim] tensors.
+        // Project, then lay out as per-head [batch*seq, head_dim] tensors.
-        //   [seq,dim] @ [dim,dim] = [seq,dim]
+        //   [B*S,dim] @ [dim,dim] = [B*S,dim]
-        //   reshape   [seq, nh, hd]
+        //   reshape   [B*S, nh, hd]
         //   qk-norm   per-head RMSNorm over hd (Qwen3-style; Q/K only, before RoPE)
-        //   rope      (kernel expects exactly [tokens, heads, head_dim])
+        //   rope      [B*S, nh, hd] with per-sequence position (period = seq)
-        //   transpose [nh, seq, hd]  → split into nh × [seq, hd]
+        //   transpose [nh, B*S, hd]  → split into nh × [B*S, hd]
         let to_heads = |proj: Var, norm: Option<&Var>| -> Vec<Var> {
-            let r = ops::reshape(&proj, &[seq, nh, hd]);
+            let r = ops::reshape(&proj, &[total, nh, hd]);
             let r = match norm {
-                // Per-head RMSNorm: flatten the (seq,nh) head rows, norm over hd,
+                // Per-head RMSNorm: flatten the (B*S,nh) head rows, norm over hd,
                 // restore. RoPE follows on the normed Q/K (mirrors xserv qwen3.rs).
                 Some(gamma) => {
-                    let flat = ops::reshape(&r, &[seq * nh, hd]);
+                    let flat = ops::reshape(&r, &[total * nh, hd]);
                     let normed = ops::rms_norm(&flat, gamma, self.cfg.eps);
-                    let r = ops::reshape(&normed, &[seq, nh, hd]);
+                    let r = ops::reshape(&normed, &[total, nh, hd]);
-                    ops::rope(&r, self.cfg.rope_theta)
+                    ops::rope(&r, self.cfg.rope_theta, seq)
                 }
                 None => r,
             };
-            let t = ops::transpose_3d01(&r); // [nh, seq, hd]
+            let t = ops::transpose_3d01(&r); // [nh, B*S, hd]
-            ops::split_heads(&t)
+            ops::split_heads(&t) // nh × [B*S, hd]
         };
 
         let q = to_heads(ops::matmul(x, &b.wq), Some(&b.q_norm));
         let k = to_heads(ops::matmul(x, &b.wk), Some(&b.k_norm));
         let v = to_heads(ops::matmul(x, &b.wv), None);
 
-        // Per-head scaled-dot-product attention with causal mask.
+        // Per-head SDPA. Within a head, attention must NOT cross sequences, so we
-        let heads_out: Vec<Var> = (0..nh)
+        // split the [B*S, hd] head into its B per-sequence [seq, hd] blocks
-            .map(|i| {
+        // (reshape to [B,seq,hd] + split over the batch axis), run masked SDPA on
-                let kt = ops::transpose_2d(&k[i]); // [hd, seq]
+        // each, then reassemble. The big GEMMs above stay batched; this inner
-                let scores = ops::scale(&ops::matmul(&q[i], &kt), scale); // [seq,seq]
+        // attention is the only sequence-aware (looped) part. Backward falls out of
+        // the matmul/softmax/scale/split/merge nodes — no new gradient to verify.
+        let attend_seq = |qh: &Var, kh: &Var, vh: &Var| -> Var {
+            let kt = ops::transpose_2d(kh); // [hd, seq]
+            let scores = ops::scale(&ops::matmul(qh, &kt), scale); // [seq,seq]
             let scores = ops::add(&scores, mask); // causal
             let probs = ops::softmax(&scores);
-                ops::matmul(&probs, &v[i]) // [seq, hd]
+            ops::matmul(&probs, vh) // [seq, hd]
+        };
+
+        let heads_out: Vec<Var> = (0..nh)
+            .map(|i| {
+                if batch == 1 {
+                    return attend_seq(&q[i], &k[i], &v[i]); // [seq, hd]
+                }
+                // [B*S, hd] → [B, seq, hd] → B × [seq, hd]
+                let qb = ops::split_heads(&ops::reshape(&q[i], &[batch, seq, hd]));
+                let kb = ops::split_heads(&ops::reshape(&k[i], &[batch, seq, hd]));
+                let vb = ops::split_heads(&ops::reshape(&v[i], &[batch, seq, hd]));
+                let per_seq: Vec<Var> = (0..batch)
+                    .map(|s| attend_seq(&qb[s], &kb[s], &vb[s]))
+                    .collect();
+                // B × [seq, hd] → [B, seq, hd] → [B*S, hd]
+                let merged = ops::merge_heads(&per_seq);
+                ops::reshape(&merged, &[total, hd])
             })
             .collect();
 
-        // Stack heads back: nh × [seq,hd] → [nh,seq,hd] → [seq,nh,hd] → [seq,dim].
+        // Stack heads back: nh × [B*S,hd] → [nh,B*S,hd] → [B*S,nh,hd] → [B*S,dim].
-        let merged = ops::merge_heads(&heads_out); // [nh, seq, hd]
+        let merged = ops::merge_heads(&heads_out); // [nh, B*S, hd]
-        let t = ops::transpose_3d01(&merged); // [seq, nh, hd]
+        let t = ops::transpose_3d01(&merged); // [B*S, nh, hd]
-        let concat = ops::reshape(&t, &[seq, nh * hd]); // [seq, dim]
+        let concat = ops::reshape(&t, &[total, nh * hd]); // [B*S, dim]
         ops::matmul(&concat, &b.wo) // out projection
     }
 
-    /// SwiGLU MLP: `down( silu(gate(x)) ∘ up(x) )`. `x`:[seq,dim].
+    /// SwiGLU MLP: `down( silu(gate(x)) ∘ up(x) )`. `x`:[batch*seq,dim].
     fn swiglu_mlp(&self, b: &Block, x: &Var) -> Var {
         let gate = ops::matmul(x, &b.w_gate); // [seq, ffn_hidden]
         let up = ops::matmul(x, &b.w_up); // [seq, ffn_hidden]
@@ -216,3 +271,17 @@ pub fn param_to_host(v: &Var) -> Vec<f32> {
 pub fn ids_tensor(ids: &[i32], device: Device) -> Tensor {
     Tensor::from_slice(ids, &[ids.len()]).to_device(device)
 }
+
+/// Flatten `batch` equal-length sequences into one `[batch*seq]` I32 tensor in
+/// sequence-major order (the layout `forward_batched` expects). Each row of
+/// `seqs` is one sequence; all must have the same length.
+pub fn batched_ids_tensor(seqs: &[Vec<i32>], device: Device) -> Tensor {
+    assert!(!seqs.is_empty(), "empty batch");
+    let seq = seqs[0].len();
+    let mut flat = Vec::with_capacity(seqs.len() * seq);
+    for s in seqs {
+        assert_eq!(s.len(), seq, "ragged batch: sequences must be equal length");
+        flat.extend_from_slice(s);
+    }
+    Tensor::from_slice(&flat, &[flat.len()]).to_device(device)
+}

crates/xtrain-model/tests/batched.rs

@@ -0,0 +1,142 @@
+// T10 batched-forward equivalence: a batched forward over B sequences must equal
+// the old single-sequence path (run each sequence on its own, concatenate the
+// logits) — both for the forward logits AND every parameter's gradient.
+//
+// This is THE on-GPU correctness gate for batching (no PyTorch needed): if the
+// per-sequence RoPE position, per-sequence causal masking, or any flattened op
+// were wrong, the batched logits/grads would drift from the looped reference.
+//
+// Forward equivalence: batched logits[b*S+i] == single-seq-b logits[i].
+// Gradient equivalence: the batched loss is the mean over all B*S rows, i.e.
+// (1/B)·Σ_b mean_i(loss_b); summing the B single-sequence losses and scaling by
+// 1/B gives the SAME scalar, so their summed grads (tape fan-out) ×1/B match the
+// batched grads. We check that.
+#![cfg(not(no_cuda))]
+
+use xtrain_cuda::device;
+use xtrain_model::{Config, TinyTransformer, batched_ids_tensor, ids_tensor};
+use xtrain_tensor::Device;
+
+fn fill(n: usize, seed: u64, scale: f32) -> Vec<f32> {
+    let mut state = seed
+        .wrapping_mul(2862933555777941757)
+        .wrapping_add(3037000493);
+    (0..n)
+        .map(|_| {
+            state = state
+                .wrapping_mul(6364136223846793005)
+                .wrapping_add(1442695040888963407);
+            (((state >> 33) as f32 / (1u64 << 31) as f32) - 0.5) * 2.0 * scale
+        })
+        .collect()
+}
+
+fn build(cfg: Config, device: Device) -> TinyTransformer {
+    let mut seed = 1u64;
+    TinyTransformer::new(cfg, device, |shape| {
+        seed = seed.wrapping_add(1);
+        let n: usize = shape.iter().product();
+        if shape.len() == 1 {
+            fill(n, seed, 0.02).iter().map(|v| v + 1.0).collect()
+        } else {
+            fill(n, seed, 0.08)
+        }
+    })
+}
+
+fn host(t: &xtrain_tensor::Tensor) -> Vec<f32> {
+    t.to_device(Device::Cpu).as_slice::<f32>().to_vec()
+}
+
+#[test]
+fn batched_matches_looped_single_sequence() {
+    assert!(device::device_count().unwrap() > 0, "no CUDA device");
+    device::set_device(0).unwrap();
+    let device = Device::Cuda(0);
+
+    let mut cfg = Config::tiny();
+    cfg.vocab = 16;
+    let batch = 3usize;
+    let seq = 5usize;
+    // B distinct sequences (sequence-major), within vocab.
+    let seqs: Vec<Vec<i32>> = (0..batch)
+        .map(|b| {
+            (0..seq)
+                .map(|i| ((b * 7 + i * 3 + 1) % cfg.vocab) as i32)
+                .collect()
+        })
+        .collect();
+    let tgts: Vec<Vec<i32>> = (0..batch)
+        .map(|b| {
+            (0..seq)
+                .map(|i| ((b * 5 + i * 2 + 2) % cfg.vocab) as i32)
+                .collect()
+        })
+        .collect();
+
+    // --- Batched forward: ONE pass over [B*S]. ---
+    let bmodel = build(cfg, device);
+    let bids = batched_ids_tensor(&seqs, device);
+    let blogits = host(&bmodel.forward_batched(&bids, batch).value());
+
+    // --- Looped reference: each sequence on its own, concatenate logits. ---
+    let smodel = build(cfg, device);
+    let mut slogits = Vec::with_capacity(batch * seq * cfg.vocab);
+    for s in &seqs {
+        let ids = ids_tensor(s, device);
+        slogits.extend(host(&smodel.forward(&ids).value()));
+    }
+
+    // Forward equivalence (fp GEMM rounding only differs in summation order).
+    let max_rel = blogits
+        .iter()
+        .zip(&slogits)
+        .map(|(b, s)| (b - s).abs() / s.abs().max(1e-4))
+        .fold(0.0f32, f32::max);
+    println!("batched vs looped: logits max rel err = {max_rel:.3e}");
+    assert!(max_rel < 1e-3, "batched logits diverged: {max_rel:.3e}");
+
+    // --- Gradient equivalence. ---
+    // Batched: loss = mean over B*S rows; one backward.
+    let bparams = bmodel.params();
+    let btgt = batched_ids_tensor(&tgts, device);
+    let bloss = bmodel.loss_batched(&bids, &btgt, batch);
+    let bloss_val = host(&bloss.value())[0];
+    bloss.backward();
+
+    // Looped: Σ_b loss_b (each a per-sequence mean), then grad ×(1/B) == batched.
+    let sparams = smodel.params();
+    let mut sloss_sum = 0.0f32;
+    for (s, t) in seqs.iter().zip(&tgts) {
+        let ids = ids_tensor(s, device);
+        let tg = ids_tensor(t, device);
+        let l = smodel.loss(&ids, &tg);
+        sloss_sum += host(&l.value())[0];
+        l.backward();
+    }
+    println!(
+        "batched loss = {bloss_val:.6}  looped mean = {:.6}",
+        sloss_sum / batch as f32
+    );
+    assert!(
+        (bloss_val - sloss_sum / batch as f32).abs() < 1e-4,
+        "batched loss != looped mean"
+    );
+
+    let mut max_grad_rel = 0.0f32;
+    for (bp, sp) in bparams.iter().zip(&sparams) {
+        let bg = host(&bp.grad().expect("batched grad"));
+        let sg = host(&sp.grad().expect("looped grad"));
+        for (g_b, g_s) in bg.iter().zip(&sg) {
+            // looped grad is the SUM over B sequences; ×(1/B) recovers the mean.
+            let g_s = g_s / batch as f32;
+            let rel = (g_b - g_s).abs() / g_s.abs().max(1e-4);
+            max_grad_rel = max_grad_rel.max(rel);
+        }
+    }
+    println!("batched vs looped: grad max rel err = {max_grad_rel:.3e}");
+    assert!(
+        max_grad_rel < 5e-3,
+        "batched grads diverged: {max_grad_rel:.3e}"
+    );
+}

crates/xtrain-model/tests/parity.py

@@ -55,10 +55,13 @@ NH = int(cfg["n_heads"])
 HD = int(cfg["head_dim"])
 EPS = float(cfg["eps"])
 THETA = float(cfg["rope_theta"])
+# Batched: B sequences of length SEQ, flattened sequence-major to [B*SEQ] ids.
+B = int(cfg.get("batch", "1"))
+SEQ = int(cfg["seq"])
 
 ids = read_ids("ids.txt")
 targets = read_ids("targets.txt")
-SEQ = len(ids)
+assert len(ids) == B * SEQ, f"ids {len(ids)} != B*SEQ {B*SEQ}"
 
 # Load params as leaf tensors requiring grad (float64 for a clean reference).
 P = {}
@@ -76,15 +79,16 @@ def rms_norm(x, gamma):
     return x * torch.rsqrt(ms + EPS) * gamma
 
 
-def rope(x):  # x: [seq, nh, hd], position = token index, matching the kernel
+def rope(x):  # x: [B*SEQ, nh, hd], position = (row % SEQ) — resets per sequence
     half = HD // 2
     out = torch.empty_like(x)
     i = torch.arange(half, dtype=torch.float64)
     freq = THETA ** (-(2.0 * i) / HD)  # [half]
-    pos = torch.arange(SEQ, dtype=torch.float64).reshape(SEQ, 1)  # [seq,1]
+    # Position within each sequence: rows 0..SEQ for seq 0, 0..SEQ for seq 1, ...
-    ang = pos * freq  # [seq, half]
+    pos = (torch.arange(B * SEQ, dtype=torch.float64) % SEQ).reshape(B * SEQ, 1)
-    c = torch.cos(ang).reshape(SEQ, 1, half)
+    ang = pos * freq  # [B*SEQ, half]
-    s = torch.sin(ang).reshape(SEQ, 1, half)
+    c = torch.cos(ang).reshape(B * SEQ, 1, half)
+    s = torch.sin(ang).reshape(B * SEQ, 1, half)
     x0 = x[..., :half]
     x1 = x[..., half:]
     out[..., :half] = x0 * c - x1 * s
@@ -102,26 +106,30 @@ for l in range(NL):
                     "ffn_norm", "w_gate", "w_up", "w_down"]})
 
 idx = torch.tensor(ids, dtype=torch.long)
+# Per-sequence causal mask (broadcast over the batch); NO cross-sequence attention.
 mask = torch.triu(torch.full((SEQ, SEQ), -1.0e9, dtype=torch.float64), diagonal=1)
 
-h = emb[idx]  # [seq, dim]
+h = emb[idx]  # [B*SEQ, dim]  (everything stays flattened, matching the Rust path)
 for L in layers:
     # Attention
     x = rms_norm(h, L["attn_norm"])
-    q = (x @ L["wq"]).reshape(SEQ, NH, HD)
+    q = (x @ L["wq"]).reshape(B * SEQ, NH, HD)
-    k = (x @ L["wk"]).reshape(SEQ, NH, HD)
+    k = (x @ L["wk"]).reshape(B * SEQ, NH, HD)
-    v = (x @ L["wv"]).reshape(SEQ, NH, HD)
+    v = (x @ L["wv"]).reshape(B * SEQ, NH, HD)
     # Per-head QK-norm (Qwen3-style), before RoPE.
     q = rms_norm(q, L["q_norm"])
     k = rms_norm(k, L["k_norm"])
-    q = rope(q).transpose(0, 1)  # [nh, seq, hd]
+    q = rope(q)  # [B*SEQ, nh, hd]
-    k = rope(k).transpose(0, 1)
+    k = rope(k)
-    v = v.transpose(0, 1)
+    # Reshape to [B, NH, SEQ, HD] so attention runs within each sequence.
+    q = q.reshape(B, SEQ, NH, HD).transpose(1, 2)  # [B, nh, seq, hd]
+    k = k.reshape(B, SEQ, NH, HD).transpose(1, 2)
+    v = v.reshape(B, SEQ, NH, HD).transpose(1, 2)
     scale = 1.0 / math.sqrt(HD)
-    scores = (q @ k.transpose(-1, -2)) * scale + mask  # [nh, seq, seq]
+    scores = (q @ k.transpose(-1, -2)) * scale + mask  # [B, nh, seq, seq]
     probs = torch.softmax(scores, dim=-1)
-    out = probs @ v  # [nh, seq, hd]
+    out = probs @ v  # [B, nh, seq, hd]
-    out = out.transpose(0, 1).reshape(SEQ, DIM)  # [seq, dim]
+    out = out.transpose(1, 2).reshape(B * SEQ, DIM)  # [B*SEQ, dim]
     attn = out @ L["wo"]
     h = h + attn
     # MLP
@@ -133,7 +141,7 @@ for L in layers:
     h = h + mlp
 
 h = rms_norm(h, final_norm)
-logits = h @ lm_head  # [seq, vocab]
+logits = h @ lm_head  # [B*SEQ, vocab]
 
 loss = torch.nn.functional.cross_entropy(
     logits, torch.tensor(targets, dtype=torch.long), reduction="mean")

crates/xtrain-model/tests/parity_dump.rs

@@ -53,12 +53,17 @@ fn dump_for_parity() {
     );
     fs::create_dir_all(&dir).unwrap();
 
-    // Fixed config + ids (independent of any text, for reproducibility).
+    // Fixed config + ids (independent of any text, for reproducibility). B>1 so
+    // the batched forward is exercised: 2 sequences of length 4, flattened
+    // sequence-major to [B*S]=8 ids. Per-sequence RoPE position (resets at the
+    // sequence boundary) + per-sequence causal masking (no cross-sequence
+    // attention) are both checked against PyTorch.
     let mut cfg = Config::tiny();
     cfg.vocab = 12;
-    let ids: Vec<i32> = vec![3, 1, 4, 1, 5, 9, 2, 6];
+    let batch = 2usize;
+    let seq = 4usize;
+    let ids: Vec<i32> = vec![3, 1, 4, 1, 5, 9, 2, 6]; // [B*S], sequence-major
     let targets: Vec<i32> = vec![1, 4, 1, 5, 9, 2, 6, 0];
-    let seq = ids.len();
 
     // Same deterministic init as the overfit test.
     let mut seed = 1u64;
@@ -83,6 +88,7 @@ fn dump_for_parity() {
         writeln!(f, "ffn_hidden {}", cfg.ffn_hidden).unwrap();
         writeln!(f, "eps {:e}", cfg.eps).unwrap();
         writeln!(f, "rope_theta {:e}", cfg.rope_theta).unwrap();
+        writeln!(f, "batch {batch}").unwrap();
         writeln!(f, "seq {seq}").unwrap();
     }
     {
@@ -105,10 +111,11 @@ fn dump_for_parity() {
         write_vec(&dir, &format!("w_{name}.txt"), &param_to_host(p), &shape);
     }
 
-    // Forward logits + loss, then backward → per-param grads.
+    // Batched forward logits + loss (B sequences as one forward), then backward
+    // → per-param grads.
     let ids_t = ids_tensor(&ids, device);
     let targets_t = ids_tensor(&targets, device);
-    let logits = model.forward(&ids_t);
+    let logits = model.forward_batched(&ids_t, batch);
     write_vec(
         &dir,
         "logits.txt",
@@ -116,7 +123,7 @@ fn dump_for_parity() {
         logits.value().shape(),
     );
 
-    let loss = model.loss(&ids_t, &targets_t);
+    let loss = model.loss_batched(&ids_t, &targets_t, batch);
     let loss_val = param_to_host(&loss)[0];
     {
         let mut f = fs::File::create(dir.join("loss.txt")).unwrap();

crates/xtrain-tensor/src/tensor.rs

@@ -454,13 +454,20 @@ impl Tensor {
         dx
     }
 
-    /// RoPE forward (rotate_half). `self`:[tokens,heads,head_dim]; the position
+    /// RoPE forward (rotate_half). `self`:[tokens,heads,head_dim]; each token's
-    /// of each token is its row index. Returns the rotated tensor.
+    /// position is `row % period`. `period` = sequence length, so a flattened
+    /// batch `[B*S,heads,head_dim]` gets per-sequence positions (pass `period=S`);
+    /// pass `period=tokens` for a single sequence (position = row). Returns the
+    /// rotated tensor.
     #[cfg(not(no_cuda))]
-    pub fn rope(&self, theta: f32) -> Self {
+    pub fn rope(&self, theta: f32, period: usize) -> Self {
         assert_eq!(self.ndim(), 3, "rope requires [tokens,heads,head_dim]");
         let (tokens, heads, head_dim) = (self.shape[0], self.shape[1], self.shape[2]);
         assert_eq!(head_dim % 2, 0, "head_dim must be even");
+        assert!(
+            period > 0 && tokens % period == 0,
+            "tokens must be a multiple of period"
+        );
         let out = Tensor::zeros(&self.shape, DType::F32, self.device());
         unsafe {
             xtrain_cuda::ffi::launch_rope_f32(
@@ -470,6 +477,7 @@ impl Tensor {
                 heads as i32,
                 head_dim as i32,
                 theta,
+                period as i32,
                 std::ptr::null_mut(),
             );
         }
@@ -477,9 +485,9 @@ impl Tensor {
     }
 
     /// RoPE backward: apply the inverse (transpose) rotation to `dy`. RoPE is an
-    /// orthogonal map, so it needs no cached forward values, only `theta`.
+    /// orthogonal map, so it needs no cached forward values, only `theta`/`period`.
     #[cfg(not(no_cuda))]
-    pub fn rope_backward(dy: &Tensor, theta: f32) -> Self {
+    pub fn rope_backward(dy: &Tensor, theta: f32, period: usize) -> Self {
         let (tokens, heads, head_dim) = (dy.shape[0], dy.shape[1], dy.shape[2]);
         let dx = Tensor::zeros(&dy.shape, DType::F32, dy.device());
         unsafe {
@@ -490,6 +498,7 @@ impl Tensor {
                 heads as i32,
                 head_dim as i32,
                 theta,
+                period as i32,
                 std::ptr::null_mut(),
             );
         }

crates/xtrain-train/src/train_loop.rs

@@ -1,18 +1,20 @@
-//! The training loop: sample sequences → forward `loss` → backward → grad clip
+//! The training loop: sample a batch of sequences → ONE batched forward `loss` →
-//! (with batch averaging) → AdamW step → zero grads; with an LR schedule,
+//! backward → grad clip → AdamW step → zero grads; with an LR schedule, periodic
-//! periodic loss logging, and periodic checkpointing.
+//! loss logging, and periodic checkpointing.
 //!
-//! The T5 model is single-sequence, so a "batch" of `batch_size` sequences is
+//! Since T10 the model is batched (`loss_batched`): `batch_size` sequences are
-//! handled by running forward+backward on each and letting the tape SUM their
+//! flattened to `[batch*seq]` and run as a SINGLE forward/backward, so the linear
-//! grads (its fan-out rule); the clip pass then multiplies by `1/batch_size` to
+//! projections become big `[batch*seq, dim]` GEMMs that fill the GPU. The
-//! recover the batch-mean gradient before clipping + the optimizer step.
+//! cross-entropy mean is over all `batch*seq` rows — already the batch-mean loss,
+//! so backward yields the batch-mean gradient directly (clip pre-scale = 1.0; no
+//! more "loop B times + SUM + ×1/batch" hack).
 
 #![cfg(not(no_cuda))]
 
 use std::path::PathBuf;
 use std::time::Instant;
 
-use xtrain_model::{TinyTransformer, ids_tensor};
+use xtrain_model::{TinyTransformer, batched_ids_tensor, ids_tensor};
 use xtrain_optim::GpuAdamW;
 use xtrain_tensor::Device;
 
@@ -67,7 +69,6 @@ pub fn train(
     let mut losses = Vec::with_capacity(cfg.steps);
     let mut evals = Vec::new();
     let mut best_val: Option<f32> = None;
-    let inv_batch = 1.0 / cfg.batch_size as f32;
     let start = Instant::now();
     let mut tokens_seen: u64 = 0;
     // Best-val checkpointing only kicks in when we actually evaluate.
@@ -76,22 +77,26 @@ pub fn train(
     for step in 0..cfg.steps {
         let lr = cfg.schedule.lr(step);
 
-        // Accumulate grads over `batch_size` sequences (tape SUMs them).
+        // Sample `batch_size` sequences and run them as ONE batched forward/
-        let mut step_loss = 0.0f32;
+        // backward. The CE mean over all batch*seq rows is the batch-mean loss, so
+        // backward already yields the batch-mean gradient (clip pre-scale = 1.0).
+        let mut inputs = Vec::with_capacity(cfg.batch_size);
+        let mut targets_v = Vec::with_capacity(cfg.batch_size);
         for _ in 0..cfg.batch_size {
             let (input, target) = corpus.sample(cfg.seq_len, &mut rng);
-            let ids = ids_tensor(&input, device);
+            inputs.push(input);
-            let targets = ids_tensor(&target, device);
+            targets_v.push(target);
-            let loss = model.loss(&ids, &targets);
-            step_loss += read_scalar(&loss);
-            loss.backward();
-            tokens_seen += cfg.seq_len as u64;
         }
-        step_loss *= inv_batch;
+        let ids = batched_ids_tensor(&inputs, device);
+        let targets = batched_ids_tensor(&targets_v, device);
+        let loss = model.loss_batched(&ids, &targets, cfg.batch_size);
+        let step_loss = read_scalar(&loss);
+        loss.backward();
+        tokens_seen += (cfg.batch_size * cfg.seq_len) as u64;
         losses.push(step_loss);
 
-        // Average the summed grads (×1/batch) and clip to the global norm.
+        // Backward already produced the batch-mean gradient — just clip it.
-        let gnorm = clip_grad_norm_gpu(&params, cfg.max_grad_norm, inv_batch);
+        let gnorm = clip_grad_norm_gpu(&params, cfg.max_grad_norm, 1.0);
         opt.step(lr, &params);
         for p in &params {
             p.zero_grad();

csrc/ops/nn.cu

@@ -215,14 +215,20 @@ void launch_silu_dx_f32(const float* x, const float* dy, float* dx, int n, void*
 // dx[i+h] = dy[i+h]*cos - dy[i]*sin
 // =====================================================================
 
-__global__ void rope_k(const float* x, float* y, int heads, int head_dim, float theta) {
+// `period` is the sequence length: a flattened batch lays B sequences end to end
+// along the `tokens` axis, so each token's RoPE position is its index WITHIN its
+// own sequence, `tok % period`. With period == tokens (single sequence) this is
+// the original position = row.
+__global__ void rope_k(const float* x, float* y, int heads, int head_dim,
+                       float theta, int period) {
     int tok = blockIdx.x;
     int head = blockIdx.y;
     int half = head_dim / 2;
     int i = threadIdx.x;
     if (i >= half) return;
+    int pos = tok % period;
     float freq = powf(theta, -(float)(2 * i) / (float)head_dim);
-    float angle = (float)tok * freq;
+    float angle = (float)pos * freq;
     float c = cosf(angle), sn = sinf(angle);
     int base = (tok * heads + head) * head_dim;
     float x0 = x[base + i], x1 = x[base + i + half];
@@ -230,20 +236,22 @@ __global__ void rope_k(const float* x, float* y, int heads, int head_dim, float
     y[base + i + half] = x1 * c + x0 * sn;
 }
 void launch_rope_f32(const float* x, float* y, int tokens, int heads,
-                     int head_dim, float theta, void* s) {
+                     int head_dim, float theta, int period, void* s) {
     dim3 grid(tokens, heads);
     int blk = head_dim / 2;
-    rope_k<<<grid, blk, 0, (cudaStream_t)s>>>(x, y, heads, head_dim, theta);
+    rope_k<<<grid, blk, 0, (cudaStream_t)s>>>(x, y, heads, head_dim, theta, period);
 }
 
-__global__ void rope_dx_k(const float* dy, float* dx, int heads, int head_dim, float theta) {
+__global__ void rope_dx_k(const float* dy, float* dx, int heads, int head_dim,
+                         float theta, int period) {
     int tok = blockIdx.x;
     int head = blockIdx.y;
     int half = head_dim / 2;
     int i = threadIdx.x;
     if (i >= half) return;
+    int pos = tok % period;
     float freq = powf(theta, -(float)(2 * i) / (float)head_dim);
-    float angle = (float)tok * freq;
+    float angle = (float)pos * freq;
     float c = cosf(angle), sn = sinf(angle);
     int base = (tok * heads + head) * head_dim;
     float d0 = dy[base + i], d1 = dy[base + i + half];
@@ -251,10 +259,10 @@ __global__ void rope_dx_k(const float* dy, float* dx, int heads, int head_dim, f
     dx[base + i + half] = d1 * c - d0 * sn;
 }
 void launch_rope_dx_f32(const float* dy, float* dx, int tokens, int heads,
-                        int head_dim, float theta, void* s) {
+                        int head_dim, float theta, int period, void* s) {
     dim3 grid(tokens, heads);
     int blk = head_dim / 2;
-    rope_dx_k<<<grid, blk, 0, (cudaStream_t)s>>>(dy, dx, heads, head_dim, theta);
+    rope_dx_k<<<grid, blk, 0, (cudaStream_t)s>>>(dy, dx, heads, head_dim, theta, period);
 }
 
 // =====================================================================