#!/usr/bin/env python3
"""Single-shot DSV4 inference — 8-GPU pipeline parallel with mHC.

Loads the full NVFP4 checkpoint across 8 B200 GPUs. Includes:
- mHC (Manifold-Constrained Hyper-Connections) — load-bearing residual
- Q low-rank projection + KV projection
- RoPE (partial, last 64 dims)
- Production FMHA kernel (tcgen05 MMA, hd=512, 128 heads)
- Output projection: wo_a (grouped BMM) → wo_b (NVFP4)
- Shared expert FFN (SwiGLU)
- NVFP4 dequant → BF16 matmul baseline for linear layers

Missing (causing incorrect output):
- Routed MoE experts (384 experts, top-6)
- KV cache across decode steps
- Compressor + indexer (CSA/HCA compressed KV)

Usage (on B200):
    source /root/dsv4-nvfp4-workspace/venv/bin/activate
    cd /root/dsv4-nvfp4-workspace/kernel
    python3 single_shot_inference.py
"""
import os, sys, time, json, math
import torch
from pathlib import Path

CHECKPOINT_DIR = "/root/nvidia-meeting/DeepSeek-V4-Pro-NVFP4"
MAX_NEW_TOKENS = 10
PROMPT = "The capital of France is"
NUM_GPUS = 8

# =====================================================================
# NVFP4 dequantization
# =====================================================================

FP4_LUT = torch.tensor([0., 2., 3., 4., 6., 8., 12., 24.])

def dequant_nvfp4_weight(weight, weight_scale, weight_scale_2):
    out_dim = weight.shape[0]
    in_packed = weight.shape[1]
    in_features = in_packed * 2
    low = (weight & 0x0F).to(torch.int8)
    high = (weight >> 4).to(torch.int8)
    low_sign, low_idx = (low >> 3).bool(), (low & 0x07).long()
    high_sign, high_idx = (high >> 3).bool(), (high & 0x07).long()
    lut = FP4_LUT.to(device=weight.device, dtype=torch.float32)
    low_f = lut[low_idx] * torch.where(low_sign, -1.0, 1.0)
    high_f = lut[high_idx] * torch.where(high_sign, -1.0, 1.0)
    w_f = torch.stack([low_f, high_f], dim=-1).reshape(out_dim, in_features)
    scale_f = weight_scale.float() * weight_scale_2.float()
    scale_expanded = scale_f.repeat_interleave(16, dim=1)
    return (w_f * scale_expanded).bfloat16()

def nvfp4_linear(x, weight, weight_scale, weight_scale_2):
    w = dequant_nvfp4_weight(weight, weight_scale, weight_scale_2)
    return torch.nn.functional.linear(x, w)

def bf16_linear(x, weight):
    return torch.nn.functional.linear(x, weight.bfloat16())


# =====================================================================
# mHC — Manifold-Constrained Hyper-Connections
# =====================================================================

def sinkhorn_knopp(M, t_max=20, eps=1e-6):
    """Project (T, n, n) positive matrices onto Birkhoff polytope."""
    for _ in range(t_max):
        M = M / (M.sum(dim=-1, keepdim=True) + eps)
        M = M / (M.sum(dim=-2, keepdim=True) + eps)
    return M


class mHCBlock:
    """One mHC block (attention or FFN).
    
    Checkpoint weight mapping:
      fn:    (24, 28672) FP32 = stacked [W_pre(4,K); W_res(16,K); W_post(4,K)]
      base:  (24,) FP32 = bias, split as [S_pre(4); S_res(16); S_post(4)]
      scale: (3,) FP32 = [alpha_pre, alpha_res, alpha_post]
    """
    def __init__(self, hidden_dim=7168, n_hc=4, t_max=20, device='cuda'):
        self.d = hidden_dim
        self.n_hc = n_hc
        self.K = n_hc * hidden_dim  # 28672
        self.t_max = t_max
        self.device = device
        self.W_stacked = None  # (24, K) FP32
        self.bias = None       # (24,) FP32
        self.alphas = None     # (3,) FP32
    
    def load_from_checkpoint(self, fn, base, scale):
        """Load from checkpoint tensors. All on target device, FP32."""
        self.W_stacked = fn.to(device=self.device, dtype=torch.float32).contiguous()
        self.bias = base.to(device=self.device, dtype=torch.float32).contiguous()
        self.alphas = scale.to(device=self.device, dtype=torch.float32).contiguous()
    
    def _dynamic_params(self, X_l):
        """Compute A_l, B_l, C_l from residual state.
        
        X_l: (T, n_hc, d) BF16
        Returns: A_l (T, n_hc), B_l (T, n_hc, n_hc) FP32, C_l (T, n_hc)
        """
        T, n, d = X_l.shape
        n_hc = self.n_hc
        
        # Flatten and project with RMSNorm
        X_flat = X_l.reshape(T, self.K)  # (T, K) BF16
        
        # RMSNorm
        x_f32 = X_flat.float()
        rms = x_f32.pow(2).mean(-1, keepdim=True).add(1e-6).rsqrt()
        x_normed = x_f32 * rms  # (T, K) FP32
        
        # Project: (T, K) @ (24, K)^T → (T, 24)
        proj = torch.nn.functional.linear(x_normed.bfloat16(), self.W_stacked.bfloat16()).float()
        proj = proj + self.bias.unsqueeze(0)  # add bias
        
        # Split into A, B, C
        i0, i1, i2 = n_hc, n_hc + n_hc * n_hc, 24
        A_raw = proj[:, :i0]        # (T, 4)
        B_raw = proj[:, i0:i1]      # (T, 16)
        C_raw = proj[:, i1:i2]      # (T, 4)
        
        # Split bias into S_pre, S_res, S_post
        S_pre = self.bias[:n_hc]
        S_res = self.bias[n_hc:n_hc + n_hc * n_hc]
        S_post = self.bias[n_hc + n_hc * n_hc:]
        
        # Apply gating + biases
        a_pre, a_res, a_post = self.alphas[0], self.alphas[1], self.alphas[2]
        A_tilde = a_pre * A_raw + S_pre.unsqueeze(0)
        B_tilde = a_res * B_raw + S_res.unsqueeze(0)
        C_tilde = a_post * C_raw + S_post.unsqueeze(0)
        
        # Constraints
        A_l = torch.sigmoid(A_tilde).bfloat16()           # (T, 4) ∈ (0,1)
        C_l = (2.0 * torch.sigmoid(C_tilde)).bfloat16()   # (T, 4) ∈ (0,2)
        B_exp = torch.exp(B_tilde).reshape(T, n_hc, n_hc) # (T, 4, 4)
        B_l = sinkhorn_knopp(B_exp, self.t_max)            # FP32, doubly stochastic
        
        return A_l, B_l, C_l
    
    def pre_block(self, X_l):
        """X_l: (T, n_hc, d) → x_in: (T, d), ctx"""
        A_l, B_l, C_l = self._dynamic_params(X_l)
        # x_in = A_l @ X_l: (T, 1, n_hc) bmm (T, n_hc, d) → (T, 1, d) → (T, d)
        x_in = torch.bmm(A_l.unsqueeze(1).float(), X_l.float()).squeeze(1).bfloat16()
        return x_in, (B_l, C_l)
    
    def post_block(self, X_l, F_out, ctx):
        """X_l: (T, n_hc, d), F_out: (T, d) → X_next: (T, n_hc, d)"""
        B_l, C_l = ctx
        # X_next = B_l @ X_l + C_l ⊗ F_out
        BX = torch.bmm(B_l, X_l.float())                    # (T, n_hc, d) FP32
        CF = C_l.unsqueeze(-1).float() * F_out.unsqueeze(1).float()  # (T, n_hc, d) FP32
        return (BX + CF).bfloat16()


# =====================================================================
# RoPE
# =====================================================================

def build_rope_cache(max_pos, head_dim, rope_dim, device, theta=10000.0):
    half = rope_dim // 2
    freqs = 1.0 / (theta ** (torch.arange(0, rope_dim, 2, dtype=torch.float32) / rope_dim))
    angles = torch.outer(torch.arange(max_pos, dtype=torch.float32), freqs)
    return torch.cos(angles).to(device), torch.sin(angles).to(device)

def apply_rope(x, positions, cos_cache, sin_cache, rope_dim):
    T, n_h, hd = x.shape
    nope = hd - rope_dim
    cos = cos_cache[positions].unsqueeze(1).to(x.dtype)
    sin = sin_cache[positions].unsqueeze(1).to(x.dtype)
    out = x.clone()
    out[:, :, nope:][..., 0::2] = x[:, :, nope:][..., 0::2] * cos - x[:, :, nope:][..., 1::2] * sin
    out[:, :, nope:][..., 1::2] = x[:, :, nope:][..., 0::2] * sin + x[:, :, nope:][..., 1::2] * cos
    return out


# =====================================================================
# Weight loading
# =====================================================================

def load_all_weights(checkpoint_dir, num_layers):
    from safetensors.torch import load_file
    
    cdir = Path(checkpoint_dir)
    index_path = cdir / "model.safetensors.index.json"
    weight_map = {}
    if index_path.exists():
        with open(index_path) as f:
            weight_map = json.load(f).get("weight_map", {})
    
    shard_names = set(weight_map.values()) if weight_map else {
        f"model-{i:05d}-of-00095.safetensors" for i in range(1, 96)
    }
    
    print(f"Loading {len(shard_names)} shards...")
    all_weights = {}
    loaded = 0
    for shard_name in sorted(shard_names):
        if not (cdir / shard_name).exists():
            continue
        data = load_file(str(cdir / shard_name))
        all_weights.update(data)
        loaded += 1
        if loaded % 20 == 0:
            print(f"  {loaded}/{len(shard_names)} shards, {len(all_weights)} tensors")
    print(f"  Done: {len(all_weights)} tensors")
    
    layer_weights = {}
    global_weights = {}
    
    print("Assigning to GPUs...")
    for key, tensor in all_weights.items():
        if key.startswith("model.layers."):
            li = int(key.split(".")[2])
            target_gpu = li % NUM_GPUS
            target_device = f"cuda:{target_gpu}"
            if li not in layer_weights:
                layer_weights[li] = {"_device": target_device, "_gpu": target_gpu}
            layer_weights[li][key] = tensor.to(target_device)
        elif key.startswith("model.embed_tokens"):
            global_weights[key] = tensor.to("cuda:0")
        elif key.startswith("model.norm"):
            global_weights[key] = tensor.to("cuda:0")
        elif key.startswith("lm_head"):
            global_weights[key] = tensor.to("cuda:0")
    
    for gpu in range(NUM_GPUS):
        alloc = torch.cuda.memory_allocated(gpu) / 1e9
        print(f"  GPU {gpu}: {alloc:.1f}GB")
    
    return layer_weights, global_weights


# =====================================================================
# Single layer forward
# =====================================================================

def forward_layer(X_l, w, li, cfg, rope_cos, rope_sin, attn_mhc, ffn_mhc):
    """Forward one layer with mHC.
    
    X_l: (1, n_hc, H) BF16 → (1, n_hc, H) BF16
    """
    device = X_l.device
    H = cfg["hidden_size"]
    n_h = cfg["num_attention_heads"]
    hd = cfg["head_dim"]
    rd = cfg["qk_rope_head_dim"]
    o_rank = cfg["o_lora_rank"]
    o_groups = cfg["o_groups"]
    n_hc = 4
    
    pre = f"model.layers.{li}.self_attn"
    T = X_l.shape[0]
    heads_per_group = n_h // o_groups
    group_input_dim = heads_per_group * hd
    
    # ==== mHC pre_block (attention) ====
    x_in, attn_ctx = attn_mhc.pre_block(X_l)  # x_in: (T, H) BF16
    
    # ==== Q projection ====
    c_Q = nvfp4_linear(x_in, w[f"{pre}.q_a_proj.weight"],
                       w[f"{pre}.q_a_proj.weight_scale"],
                       w[f"{pre}.q_a_proj.weight_scale_2"])
    q = nvfp4_linear(c_Q, w[f"{pre}.q_b_proj.weight"],
                     w[f"{pre}.q_b_proj.weight_scale"],
                     w[f"{pre}.q_b_proj.weight_scale_2"])
    
    # ==== KV projection ====
    kv = nvfp4_linear(x_in, w[f"{pre}.kv_proj.weight"],
                      w[f"{pre}.kv_proj.weight_scale"],
                      w[f"{pre}.kv_proj.weight_scale_2"])
    
    # ==== Reshape for attention ====
    q_heads = q.reshape(T, n_h, hd).permute(1, 0, 2)
    k = kv.reshape(T, 1, hd).permute(1, 0, 2)
    v = k.clone()
    
    # ==== RoPE ====
    pos = torch.tensor([0], dtype=torch.long, device=device)
    q_heads = apply_rope(q_heads, pos, rope_cos, rope_sin, rd)
    k = apply_rope(k, pos, rope_cos, rope_sin, rd)
    
    # ==== FMHA ====
    from dsv4.kernels.attention.production import dsv4_attention
    attn_out = dsv4_attention(q_heads, k, v)
    attn_out = attn_out.permute(1, 0, 2).reshape(T, n_h * hd)
    
    # ==== Output projection ====
    attn_grouped = attn_out.reshape(T, o_groups, heads_per_group, hd).reshape(T, o_groups, group_input_dim)
    oa_w = w[f"{pre}.o_a_proj.weight"].bfloat16()
    oa_3d = oa_w.reshape(o_groups, o_rank, group_input_dim)
    attn_for_bmm = attn_grouped.permute(1, 0, 2)
    grouped_out = torch.bmm(attn_for_bmm, oa_3d.transpose(1, 2))
    grouped_flat = grouped_out.permute(1, 0, 2).reshape(T, o_groups * o_rank)
    F_attn = nvfp4_linear(grouped_flat, w[f"{pre}.o_b_proj.weight"],
                           w[f"{pre}.o_b_proj.weight_scale"],
                           w[f"{pre}.o_b_proj.weight_scale_2"])
    
    # ==== mHC post_block (attention) ====
    X_l = attn_mhc.post_block(X_l, F_attn, attn_ctx)  # (T, n_hc, H)
    
    # ==== mHC pre_block (FFN) ====
    x_ffn, ffn_ctx = ffn_mhc.pre_block(X_l)
    
    # ==== FFN: shared expert ====
    se_pre = f"model.layers.{li}.mlp.shared_experts"
    se_gate_w = w.get(f"{se_pre}.gate_proj.weight")
    F_ffn = torch.zeros_like(x_ffn)
    if se_gate_w is not None:
        gate = nvfp4_linear(x_ffn, se_gate_w,
                           w[f"{se_pre}.gate_proj.weight_scale"],
                           w[f"{se_pre}.gate_proj.weight_scale_2"])
        up = nvfp4_linear(x_ffn, w[f"{se_pre}.up_proj.weight"],
                         w[f"{se_pre}.up_proj.weight_scale"],
                         w[f"{se_pre}.up_proj.weight_scale_2"])
        F_ffn = nvfp4_linear(
            torch.nn.functional.silu(gate) * up,
            w[f"{se_pre}.down_proj.weight"],
            w[f"{se_pre}.down_proj.weight_scale"],
            w[f"{se_pre}.down_proj.weight_scale_2"],
        )
    
    # ==== mHC post_block (FFN) ====
    X_l = ffn_mhc.post_block(X_l, F_ffn, ffn_ctx)
    
    return X_l


# =====================================================================
# Main
# =====================================================================

def main():
    t_start = time.time()
    print("=" * 70)
    print("DSV4 Single-Shot Inference — 8-GPU with mHC")
    print("=" * 70)
    
    with open(os.path.join(CHECKPOINT_DIR, "config.json")) as f:
        cfg = json.load(f)
    n_layers = cfg["num_hidden_layers"]
    H = cfg["hidden_size"]
    n_h = cfg["num_attention_heads"]
    hd = cfg["head_dim"]
    rd = cfg["qk_rope_head_dim"]
    n_hc = 4
    print(f"Model: {n_layers} layers, {n_h} heads, hd={hd}, rope_dim={rd}")
    
    # ==== Phase 1: Load weights ====
    print(f"\n{'='*70}\nPhase 1: Loading weights\n{'='*70}")
    layer_weights, global_weights = load_all_weights(CHECKPOINT_DIR, n_layers)
    t_loaded = time.time()
    print(f"Weight loading: {t_loaded - t_start:.1f}s")
    
    # ==== Build mHC blocks per layer ====
    print("Building mHC blocks...")
    attn_mhc_blocks = {}
    ffn_mhc_blocks = {}
    for li in range(n_layers):
        gpu = li % NUM_GPUS
        dev = f"cuda:{gpu}"
        
        # Attention mHC
        attn_fn = layer_weights[li].get(f"model.layers.{li}.attn_hc.fn")
        attn_base = layer_weights[li].get(f"model.layers.{li}.attn_hc.base")
        attn_scale = layer_weights[li].get(f"model.layers.{li}.attn_hc.scale")
        if attn_fn is not None and attn_base is not None and attn_scale is not None:
            attn_mhc = mHCBlock(hidden_dim=H, n_hc=n_hc, device=dev)
            attn_mhc.load_from_checkpoint(attn_fn, attn_base, attn_scale)
            attn_mhc_blocks[li] = attn_mhc
        
        # FFN mHC
        ffn_fn = layer_weights[li].get(f"model.layers.{li}.ffn_hc.fn")
        ffn_base = layer_weights[li].get(f"model.layers.{li}.ffn_hc.base")
        ffn_scale = layer_weights[li].get(f"model.layers.{li}.ffn_hc.scale")
        if ffn_fn is not None and ffn_base is not None and ffn_scale is not None:
            ffn_mhc = mHCBlock(hidden_dim=H, n_hc=n_hc, device=dev)
            ffn_mhc.load_from_checkpoint(ffn_fn, ffn_base, ffn_scale)
            ffn_mhc_blocks[li] = ffn_mhc
    
    print(f"  attn mHC: {len(attn_mhc_blocks)} layers")
    print(f"  ffn mHC: {len(ffn_mhc_blocks)} layers")
    
    # ==== Global weights (gpu0) ====
    torch.cuda.set_device(0)
    embed_w = global_weights.get("model.embed_tokens.weight")
    embed = torch.nn.Embedding.from_pretrained(embed_w.bfloat16())
    lm_w = global_weights.get("lm_head.weight", embed_w).bfloat16()
    final_norm_w = global_weights.get("model.norm.weight")
    
    # RoPE caches per GPU
    rope_caches = {g: build_rope_cache(8192, hd, rd, f"cuda:{g}") for g in range(NUM_GPUS)}
    
    # ==== Phase 2: Compile ====
    print(f"\n{'='*70}\nPhase 2: JIT compiling\n{'='*70}")
    from dsv4.kernels.attention.production import dsv4_attention
    dummy_q = torch.randn(n_h, 1, hd, dtype=torch.bfloat16, device='cuda:0')
    dummy_k = torch.randn(1, 1, hd, dtype=torch.bfloat16, device='cuda:0')
    try:
        _ = dsv4_attention(dummy_q, dummy_k, dummy_k.clone())
        print("  FMHA: compiled OK")
    except Exception as e:
        print(f"  FMHA error: {e}")
    t_compiled = time.time()
    print(f"Compile: {t_compiled - t_loaded:.1f}s")
    
    # ==== Phase 3: Inference ====
    print(f"\n{'='*70}\nPhase 3: Inference\n{'='*70}")
    from transformers import AutoTokenizer
    tokenizer = AutoTokenizer.from_pretrained(CHECKPOINT_DIR)
    input_ids = tokenizer.encode(PROMPT, return_tensors="pt").cuda()
    print(f"Prompt: '{PROMPT}' → {input_ids.tolist()}")
    
    generated = input_ids[0].tolist()
    
    for step in range(MAX_NEW_TOKENS):
        t0 = time.time()
        tid = torch.tensor([generated[-1]], dtype=torch.long, device='cuda:0')
        
        # Embed → mHC init state
        emb = embed(tid)  # (1, H) on gpu0
        X = emb.unsqueeze(1).expand(-1, n_hc, -1).clone()  # (1, n_hc, H)
        
        # Process layers
        for li in range(n_layers):
            gpu = li % NUM_GPUS
            target_device = f"cuda:{gpu}"
            if X.device != torch.device(target_device):
                X = X.to(target_device)
            torch.cuda.set_device(gpu)
            
            attn_mhc = attn_mhc_blocks.get(li)
            ffn_mhc = ffn_mhc_blocks.get(li)
            rc, rs = rope_caches[gpu]
            X = forward_layer(X, layer_weights[li], li, cfg, rc, rs, attn_mhc, ffn_mhc)
        
        # Back to gpu0
        X = X.to('cuda:0')
        torch.cuda.set_device(0)
        
        # Read out stream 0 → RMSNorm → lm_head
        x_out = X[:, 0, :]  # (1, H)
        if final_norm_w is not None:
            xf = x_out.float()
            rms = xf.pow(2).mean(-1, keepdim=True).add(1e-6).rsqrt()
            x_out = (xf * rms * final_norm_w.float()).bfloat16()
        
        logits = torch.nn.functional.linear(x_out, lm_w)
        next_id = torch.argmax(logits, dim=-1).item()
        generated.append(next_id)
        
        tok_str = tokenizer.decode([next_id])
        dt = time.time() - t0
        
        # Check for NaN
        has_nan = torch.isnan(logits.float()).any().item()
        logit_range = f"[{logits.float().min().item():.1f}, {logits.float().max().item():.1f}]"
        print(f"  Step {step}: {next_id} '{tok_str}' ({dt:.2f}s) logits={logit_range} nan={has_nan}")
        
        if has_nan:
            print("  NaN detected, stopping")
            break
        if next_id == tokenizer.eos_token_id:
            break
    
    out = tokenizer.decode(generated, skip_special_tokens=True)
    total = time.time() - t_start
    print(f"\n{'='*70}")
    print(f"Input:  '{PROMPT}'")
    print(f"Output: '{out}'")
    print(f"Total: {total:.1f}s (load: {t_loaded-t_start:.1f}s, compile: {t_compiled-t_loaded:.1f}s, infer: {time.time()-t_compiled:.1f}s)")
    print(f"{'='*70}")


if __name__ == "__main__":
    main()