EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Revert "[BUGFIX] Mtp torchair pd fix (#3449)" (#3500)

Browse Source 
This reverts commit b0ae203e72.

### What this PR does / why we need it?
The fix is not ready yet, conflict with #3411 need to revert first. Will
fix this issue later

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

Signed-off-by: xuyexiong <xuyexiong@huawei.com>
This commit is contained in:
xuyexiong
2025-10-17 09:42:48 +08:00
committed by GitHub
parent 3a53bbc508
commit 30e3d86b0f
1 changed files with 4 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
vllm_ascend/torchair/torchair_model_runner.py
View File

@@ -452,31 +452,10 @@ class NPUTorchairModelRunner(NPUModelRunner):
self.torchair_graph_batch_sizes.append(self.max_num_reqs)
# padded max number tokens = max_num_req * decode_token_per_req
if self.decode_token_per_req > 1:
# pd disaggregation scenario need redundant_batch_sizes to avoid each batch's seq_len exceed 16 tokens
if self.is_kv_consumer:
FIA_SEQ_LEN_LIMIT = 16
self.torchair_graph_batch_sizes = [
(graph_batch_size +
math.ceil(graph_batch_size / FIA_SEQ_LEN_LIMIT) +
math.ceil(graph_batch_size * self.decode_token_per_req /
FIA_SEQ_LEN_LIMIT / FIA_SEQ_LEN_LIMIT)) *
self.decode_token_per_req
for graph_batch_size in self.torchair_graph_batch_sizes
]
new_max_num_reqs = math.ceil(
max(self.torchair_graph_batch_sizes) /
self.decode_token_per_req)
if self.max_num_reqs < new_max_num_reqs:
logger.warning(
f"max_num_reqs is updated to {new_max_num_reqs}")
self.max_num_reqs = new_max_num_reqs
self.scheduler_config.max_num_seqs = new_max_num_reqs
else:
self.torchair_graph_batch_sizes = [
graph_batch_size * self.decode_token_per_req
for graph_batch_size in self.torchair_graph_batch_sizes
]
self.torchair_graph_batch_sizes = [
graph_batch_size * self.decode_token_per_req
for graph_batch_size in self.torchair_graph_batch_sizes
]
# NOTE: when enable_expert_parallel on A3, we need to check if `graph_batch_size` is divisible by `tp_size`
# Because we use x_active_mask for dispatch/combine op on A3, which requires that input shape should be same