| """Llama-shaped decoder run in CHUNKED PREFILL: `chunk` tokens per call, not one. |
| |
| Every other task in this family is a batch-1 decode step -- GEMV-shaped, bandwidth-bound, latency |
| regime. This one is the other half of a serving stack: a 256-token chunk goes through the whole model |
| in one call, the matmuls become real GEMMs, and the balance flips to compute. The fusion problem is the |
| same shape (one persistent kernel, all layers, no round trips) but everything about the *inside* of it |
| changes: you now have arithmetic intensity to protect rather than bandwidth to conserve. |
| |
| Only the LAST position's logits are returned, which is what a real chunked-prefill scheduler does -- |
| the intermediate positions exist only to fill the KV cache. |
| """ |
| from model import HELPERS_CORE, QUANT_FP8 |
|
|
| BODY = r''' |
| def _rope_range(x, cos, sin, pos0): |
| """x: (B, H, T, hd), positions pos0 .. pos0+T-1. Rotation in fp32, cast back.""" |
| T = x.shape[2] |
| c = cos[pos0:pos0 + T].unsqueeze(0).unsqueeze(0) # (1, 1, T, hd/2) |
| s = sin[pos0:pos0 + T].unsqueeze(0).unsqueeze(0) |
| xf = x.float() |
| x1, x2 = xf[..., ::2], xf[..., 1::2] |
| return torch.stack([x1 * c - x2 * s, x1 * s + x2 * c], dim=-1).flatten(-2).to(x.dtype) |
| |
| |
| def make_weights(cfg, seed=0, device="cuda"): |
| """Deterministic 1/sqrt(fan_in)-scaled weights. No checkpoint is shipped or downloaded.""" |
| g = torch.Generator(device=device).manual_seed(seed) |
| d, ffn, n_q, n_kv, hd = cfg["d"], cfg["ffn"], cfg["n_q"], cfg["n_kv"], cfg["hd"] |
| dt = cfg["wdtype"] |
| |
| def rnd(*shape, fan_in): |
| w = torch.randn(*shape, device=device, dtype=torch.float32, generator=g) / (fan_in ** 0.5) |
| return _quantise(w, dt) |
| |
| ones = lambda: torch.ones(d, device=device, dtype=torch.bfloat16) |
| W = {"embed": rnd(cfg["vocab"], d, fan_in=d), "final_norm": ones(), "layers": []} |
| for _ in range(cfg["layers"]): |
| W["layers"].append(dict( |
| in_norm=ones(), post_norm=ones(), |
| q=rnd(n_q * hd, d, fan_in=d), k=rnd(n_kv * hd, d, fan_in=d), |
| v=rnd(n_kv * hd, d, fan_in=d), o=rnd(d, n_q * hd, fan_in=n_q * hd), |
| gate=rnd(ffn, d, fan_in=d), up=rnd(ffn, d, fan_in=d), down=rnd(d, ffn, fan_in=ffn))) |
| return W |
| |
| |
| def make_kv(cfg, batch, prefill_len, max_seq, seed=0, device="cuda"): |
| """KV cache already holding `prefill_len` tokens. The first chunk is written at pos = prefill_len.""" |
| g = torch.Generator(device=device).manual_seed(seed + 777) |
| kv = [] |
| for _ in range(cfg["layers"]): |
| k = torch.zeros(batch, cfg["n_kv"], max_seq, cfg["hd"], device=device, dtype=torch.bfloat16) |
| v = torch.zeros_like(k) |
| k[:, :, :prefill_len] = torch.randn(batch, cfg["n_kv"], prefill_len, cfg["hd"], device=device, |
| dtype=torch.float32, generator=g).to(torch.bfloat16) * 0.5 |
| v[:, :, :prefill_len] = torch.randn(batch, cfg["n_kv"], prefill_len, cfg["hd"], device=device, |
| dtype=torch.float32, generator=g).to(torch.bfloat16) * 0.5 |
| kv.append((k, v)) |
| return kv |
| |
| |
| def make_step_args(cfg, batch, base_pos, seed, n): |
| """(token_ids, pos) per step, where token_ids is (B, chunk) and the chunk starts at pos.""" |
| g = torch.Generator(device="cuda").manual_seed(seed) |
| C = cfg["chunk"] |
| return [(torch.randint(0, cfg["vocab"], (batch, C), device="cuda", generator=g), base_pos + i * C) |
| for i in range(n)] |
| |
| |
| def build_model(weights, kv_cache, cfg, max_seq_len): |
| """UNTIMED setup. Returns whatever handle you like; the grader only passes it back to decode_step.""" |
| cos, sin = _rope_cache(cfg, max_seq_len, weights["final_norm"].device) |
| W = {"embed": _deq(weights["embed"]), "final_norm": weights["final_norm"], |
| "layers": [{k: (v if k.endswith("norm") else _deq(v)) for k, v in L.items()} |
| for L in weights["layers"]]} |
| return {"W": W, "kv": kv_cache, "cfg": cfg, "cos": cos, "sin": sin} |
| |
| |
| @torch.no_grad() |
| def decode_step(handle, token_ids, pos): |
| """One CHUNK for every sequence in the batch. Appends `chunk` positions into the cache. |
| |
| token_ids: (B, chunk) int64 pos: int, absolute position of token_ids[:, 0] |
| returns: (B, vocab) logits for the LAST position of the chunk only |
| """ |
| W, kv, cfg = handle["W"], handle["kv"], handle["cfg"] |
| cos, sin = handle["cos"], handle["sin"] |
| B, C = token_ids.shape |
| n_q, n_kv, hd = cfg["n_q"], cfg["n_kv"], cfg["hd"] |
| rep, tot = n_q // n_kv, pos + C |
| |
| # causal mask: query i (absolute position pos+i) sees keys 0 .. pos+i |
| ar = torch.arange(tot, device=token_ids.device) |
| mask = ar.unsqueeze(0) <= (pos + torch.arange(C, device=token_ids.device)).unsqueeze(1) |
| |
| x = W["embed"][token_ids] # (B, C, d) |
| for li, L in enumerate(W["layers"]): |
| h = _rms_norm(x, L["in_norm"], cfg["eps"]) |
| q = (h @ L["q"].T).view(B, C, n_q, hd).transpose(1, 2) |
| k = (h @ L["k"].T).view(B, C, n_kv, hd).transpose(1, 2) |
| v = (h @ L["v"].T).view(B, C, n_kv, hd).transpose(1, 2) |
| q = _rope_range(q, cos, sin, pos) |
| k = _rope_range(k, cos, sin, pos) |
| kc, vc = kv[li] |
| kc[:, :, pos:tot] = k |
| vc[:, :, pos:tot] = v |
| kk = kc[:, :, :tot].repeat_interleave(rep, dim=1) |
| vv = vc[:, :, :tot].repeat_interleave(rep, dim=1) |
| att = F.scaled_dot_product_attention(q, kk, vv, attn_mask=mask) |
| x = x + (att.transpose(1, 2).reshape(B, C, n_q * hd) @ L["o"].T) |
| h = _rms_norm(x, L["post_norm"], cfg["eps"]) |
| x = x + ((F.silu(h @ L["gate"].T) * (h @ L["up"].T)) @ L["down"].T) |
| x = _rms_norm(x[:, -1], W["final_norm"], cfg["eps"]) # LAST position only |
| return x @ W["embed"].T # tied lm_head |
| ''' |
|
|
| MODEL_SRC = HELPERS_CORE + QUANT_FP8 + BODY |
|
|
