dylanyunlong/project_6
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

fix(GDN): clamp gate [-5,2] + state [-65504,65504] to prevent inf/NaN

Browse Source 
Root cause from real machine test: gdn_forward.cu output abs mean = inf
- gate_raw can be positive → exp(gate) > 1 → state grows exponentially
- Over 64 tokens: exp(2.0)^64 = inf
- PyTorch ref clamps g ∈ [-5, 2] but CUDA kernel did not

Fix:
  gdn_forward.cu: clamp gate_raw ∈ [-5, 2] before exp (both kernel variants)
  gdn_forward.cu: clamp state ∈ [-65504, 65504] after update (fp16 safe range)
  qwen3_5.py: clamp g_3d before passing to SM70 kernel (belt + suspenders)
  qwen3_5.py: clamp temporal_state after decode update
This commit is contained in:
EX Engine
2026-08-10 03:38:39 +00:00
parent 388f6b2d1a
commit 8eba1750fa
2 changed files with 15 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
qwen3_6_scripts/flash_qla_sm70/csrc/gdn_forward.cu
View File

@@ -135,7 +135,7 @@ __global__ void gdn_forward_kernel(const scalar_t* __restrict__ q,
float beta_value = 0.0F;
if (threadIdx.x == 0) {
const float gate_raw = load_as_float(gate, gate_index);
gate_value = GateIsExp ? gate_raw : __expf(gate_raw);
{ const float gc = fminf(fmaxf(gate_raw, -5.0F), 2.0F); gate_value = GateIsExp ? fminf(gate_raw, 7.389F) : __expf(gc); }
beta_value = load_as_float(beta, gate_index);
}
gate_value = __shfl_sync(0xffffffffU, gate_value, 0);
@@ -191,7 +191,7 @@ __global__ void gdn_forward_kernel(const scalar_t* __restrict__ q,
for (int c = 0; c < COLS; ++c) {
const float new_state =
fmaf(k_reg[r], delta[c], gate_value * state_shard[c][r]);
state_shard[c][r] = new_state;
state_shard[c][r] = fminf(fmaxf(new_state, -65504.0F), 65504.0F);
attn_partial[c] += new_state * q_reg[r];
}
}
@@ -294,7 +294,7 @@ __global__ void gdn_forward_vlk_varlen_kernel(
float beta_value = 0.0F;
if (threadIdx.x == 0) {
const float gate_raw = load_as_float(gate, gate_index);
gate_value = GateIsExp ? gate_raw : __expf(gate_raw);
{ const float gc = fminf(fmaxf(gate_raw, -5.0F), 2.0F); gate_value = GateIsExp ? fminf(gate_raw, 7.389F) : __expf(gc); }
beta_value = load_as_float(beta, gate_index);
}
gate_value = __shfl_sync(0xffffffffU, gate_value, 0);
@@ -350,7 +350,7 @@ __global__ void gdn_forward_vlk_varlen_kernel(
for (int c = 0; c < COLS; ++c) {
const float new_state =
fmaf(k_reg[r], delta[c], gate_value * state_shard[c][r]);
state_shard[c][r] = new_state;
state_shard[c][r] = fminf(fmaxf(new_state, -65504.0F), 65504.0F);
attn_partial[c] += new_state * q_reg[r];
}
}
@@ -543,7 +543,7 @@ __global__ void gdn_decode_mixed_qkv_global_state_kernel(
for (int c = 0; c < COLS; ++c) {
const float new_state =
fmaf(k_reg[r], delta[c], gate_value * state_shard[c][r]);
state_shard[c][r] = new_state;
state_shard[c][r] = fminf(fmaxf(new_state, -65504.0F), 65504.0F);
attn_partial[c] += new_state * q_reg[r];
}
}
@@ -765,7 +765,7 @@ __global__ void gdn_decode_mixed_qkv_ddtree_state_kernel(
for (int c = 0; c < COLS; ++c) {
const float new_state =
fmaf(k_reg[r], delta[c], gate_value * state_shard[c][r]);
state_shard[c][r] = new_state;
state_shard[c][r] = fminf(fmaxf(new_state, -65504.0F), 65504.0F);
attn_partial[c] += new_state * q_reg[r];
}
}